net/netfilter/nf_conntrack_netlink.c

   1 /* Connection tracking via netlink socket. Allows for user space
   2  * protocol helpers and general trouble making from userspace.
   3  *
   4  * (C) 2001 by Jay Schulist <jschlst@samba.org>
   5  * (C) 2002-2006 by Harald Welte <laforge@gnumonks.org>
   6  * (C) 2003 by Patrick Mchardy <kaber@trash.net>
   7  * (C) 2005-2012 by Pablo Neira Ayuso <pablo@netfilter.org>
   8  *
   9  * Initial connection tracking via netlink development funded and
  10  * generally made possible by Network Robots, Inc. (www.networkrobots.com)
  11  *
  12  * Further development of this code funded by Astaro AG (http://www.astaro.com)
  13  *
  14  * This software may be used and distributed according to the terms
  15  * of the GNU General Public License, incorporated herein by reference.
  16  */
  17
  18 #include <linux/init.h>
  19 #include <linux/module.h>
  20 #include <linux/kernel.h>
  21 #include <linux/rculist.h>
  22 #include <linux/rculist_nulls.h>
  23 #include <linux/types.h>
  24 #include <linux/timer.h>
  25 #include <linux/security.h>
  26 #include <linux/skbuff.h>
  27 #include <linux/errno.h>
  28 #include <linux/netlink.h>
  29 #include <linux/spinlock.h>
  30 #include <linux/interrupt.h>
  31 #include <linux/slab.h>
  32
  33 #include <linux/netfilter.h>
  34 #include <net/netlink.h>
  35 #include <net/sock.h>
  36 #include <net/netfilter/nf_conntrack.h>
  37 #include <net/netfilter/nf_conntrack_core.h>
  38 #include <net/netfilter/nf_conntrack_expect.h>
  39 #include <net/netfilter/nf_conntrack_helper.h>
  40 #include <net/netfilter/nf_conntrack_l3proto.h>
  41 #include <net/netfilter/nf_conntrack_l4proto.h>
  42 #include <net/netfilter/nf_conntrack_tuple.h>
  43 #include <net/netfilter/nf_conntrack_acct.h>
  44 #include <net/netfilter/nf_conntrack_zones.h>
  45 #include <net/netfilter/nf_conntrack_timestamp.h>
  46 #ifdef CONFIG_NF_NAT_NEEDED
  47 #include <net/netfilter/nf_nat_core.h>
  48 #include <net/netfilter/nf_nat_l4proto.h>
  49 #include <net/netfilter/nf_nat_helper.h>
  50 #endif
  51
  52 #include <linux/netfilter/nfnetlink.h>
  53 #include <linux/netfilter/nfnetlink_conntrack.h>
  54
  55 MODULE_LICENSE("GPL");
  56
  57 static char __initdata version[] = "0.93";
  58
  59 static inline int
  60 ctnetlink_dump_tuples_proto(struct sk_buff *skb,
  61                             const struct nf_conntrack_tuple *tuple,
  62                             struct nf_conntrack_l4proto *l4proto)
  63 {
  64         int ret = 0;
  65         struct nlattr *nest_parms;
  66
  67         nest_parms = nla_nest_start(skb, CTA_TUPLE_PROTO | NLA_F_NESTED);
  68         if (!nest_parms)
  69                 goto nla_put_failure;
  70         if (nla_put_u8(skb, CTA_PROTO_NUM, tuple->dst.protonum))
  71                 goto nla_put_failure;
  72
  73         if (likely(l4proto->tuple_to_nlattr))
  74                 ret = l4proto->tuple_to_nlattr(skb, tuple);
  75
  76         nla_nest_end(skb, nest_parms);
  77
  78         return ret;
  79
  80 nla_put_failure:
  81         return -1;
  82 }
  83
  84 static inline int
  85 ctnetlink_dump_tuples_ip(struct sk_buff *skb,
  86                          const struct nf_conntrack_tuple *tuple,
  87                          struct nf_conntrack_l3proto *l3proto)
  88 {
  89         int ret = 0;
  90         struct nlattr *nest_parms;
  91
  92         nest_parms = nla_nest_start(skb, CTA_TUPLE_IP | NLA_F_NESTED);
  93         if (!nest_parms)
  94                 goto nla_put_failure;
  95
  96         if (likely(l3proto->tuple_to_nlattr))
  97                 ret = l3proto->tuple_to_nlattr(skb, tuple);
  98
  99         nla_nest_end(skb, nest_parms);
 100
 101         return ret;
 102
 103 nla_put_failure:
 104         return -1;
 105 }
 106
 107 static int
 108 ctnetlink_dump_tuples(struct sk_buff *skb,
 109                       const struct nf_conntrack_tuple *tuple)
 110 {
 111         int ret;
 112         struct nf_conntrack_l3proto *l3proto;
 113         struct nf_conntrack_l4proto *l4proto;
 114
 115         rcu_read_lock();
 116         l3proto = __nf_ct_l3proto_find(tuple->src.l3num);
 117         ret = ctnetlink_dump_tuples_ip(skb, tuple, l3proto);
 118
 119         if (ret >= 0) {
 120                 l4proto = __nf_ct_l4proto_find(tuple->src.l3num,
 121                                                tuple->dst.protonum);
 122                 ret = ctnetlink_dump_tuples_proto(skb, tuple, l4proto);
 123         }
 124         rcu_read_unlock();
 125         return ret;
 126 }
 127
 128 static inline int
 129 ctnetlink_dump_status(struct sk_buff *skb, const struct nf_conn *ct)
 130 {
 131         if (nla_put_be32(skb, CTA_STATUS, htonl(ct->status)))
 132                 goto nla_put_failure;
 133         return 0;
 134
 135 nla_put_failure:
 136         return -1;
 137 }
 138
 139 static inline int
 140 ctnetlink_dump_timeout(struct sk_buff *skb, const struct nf_conn *ct)
 141 {
 142         long timeout = ((long)ct->timeout.expires - (long)jiffies) / HZ;
 143
 144         if (timeout < 0)
 145                 timeout = 0;
 146
 147         if (nla_put_be32(skb, CTA_TIMEOUT, htonl(timeout)))
 148                 goto nla_put_failure;
 149         return 0;
 150
 151 nla_put_failure:
 152         return -1;
 153 }
 154
 155 static inline int
 156 ctnetlink_dump_protoinfo(struct sk_buff *skb, struct nf_conn *ct)
 157 {
 158         struct nf_conntrack_l4proto *l4proto;
 159         struct nlattr *nest_proto;
 160         int ret;
 161
 162         l4proto = __nf_ct_l4proto_find(nf_ct_l3num(ct), nf_ct_protonum(ct));
 163         if (!l4proto->to_nlattr)
 164                 return 0;
 165
 166         nest_proto = nla_nest_start(skb, CTA_PROTOINFO | NLA_F_NESTED);
 167         if (!nest_proto)
 168                 goto nla_put_failure;
 169
 170         ret = l4proto->to_nlattr(skb, nest_proto, ct);
 171
 172         nla_nest_end(skb, nest_proto);
 173
 174         return ret;
 175
 176 nla_put_failure:
 177         return -1;
 178 }
 179
 180 static inline int
 181 ctnetlink_dump_helpinfo(struct sk_buff *skb, const struct nf_conn *ct)
 182 {
 183         struct nlattr *nest_helper;
 184         const struct nf_conn_help *help = nfct_help(ct);
 185         struct nf_conntrack_helper *helper;
 186
 187         if (!help)
 188                 return 0;
 189
 190         helper = rcu_dereference(help->helper);
 191         if (!helper)
 192                 goto out;
 193
 194         nest_helper = nla_nest_start(skb, CTA_HELP | NLA_F_NESTED);
 195         if (!nest_helper)
 196                 goto nla_put_failure;
 197         if (nla_put_string(skb, CTA_HELP_NAME, helper->name))
 198                 goto nla_put_failure;
 199
 200         if (helper->to_nlattr)
 201                 helper->to_nlattr(skb, ct);
 202
 203         nla_nest_end(skb, nest_helper);
 204 out:
 205         return 0;
 206
 207 nla_put_failure:
 208         return -1;
 209 }
 210
 211 static int
 212 dump_counters(struct sk_buff *skb, u64 pkts, u64 bytes,
 213               enum ip_conntrack_dir dir)
 214 {
 215         enum ctattr_type type = dir ? CTA_COUNTERS_REPLY: CTA_COUNTERS_ORIG;
 216         struct nlattr *nest_count;
 217
 218         nest_count = nla_nest_start(skb, type | NLA_F_NESTED);
 219         if (!nest_count)
 220                 goto nla_put_failure;
 221
 222         if (nla_put_be64(skb, CTA_COUNTERS_PACKETS, cpu_to_be64(pkts)) ||
 223             nla_put_be64(skb, CTA_COUNTERS_BYTES, cpu_to_be64(bytes)))
 224                 goto nla_put_failure;
 225
 226         nla_nest_end(skb, nest_count);
 227
 228         return 0;
 229
 230 nla_put_failure:
 231         return -1;
 232 }
 233
 234 static int
 235 ctnetlink_dump_counters(struct sk_buff *skb, const struct nf_conn *ct,
 236                         enum ip_conntrack_dir dir, int type)
 237 {
 238         struct nf_conn_counter *acct;
 239         u64 pkts, bytes;
 240
 241         acct = nf_conn_acct_find(ct);
 242         if (!acct)
 243                 return 0;
 244
 245         if (type == IPCTNL_MSG_CT_GET_CTRZERO) {
 246                 pkts = atomic64_xchg(&acct[dir].packets, 0);
 247                 bytes = atomic64_xchg(&acct[dir].bytes, 0);
 248         } else {
 249                 pkts = atomic64_read(&acct[dir].packets);
 250                 bytes = atomic64_read(&acct[dir].bytes);
 251         }
 252         return dump_counters(skb, pkts, bytes, dir);
 253 }
 254
 255 static int
 256 ctnetlink_dump_timestamp(struct sk_buff *skb, const struct nf_conn *ct)
 257 {
 258         struct nlattr *nest_count;
 259         const struct nf_conn_tstamp *tstamp;
 260
 261         tstamp = nf_conn_tstamp_find(ct);
 262         if (!tstamp)
 263                 return 0;
 264
 265         nest_count = nla_nest_start(skb, CTA_TIMESTAMP | NLA_F_NESTED);
 266         if (!nest_count)
 267                 goto nla_put_failure;
 268
 269         if (nla_put_be64(skb, CTA_TIMESTAMP_START, cpu_to_be64(tstamp->start)) ||
 270             (tstamp->stop != 0 && nla_put_be64(skb, CTA_TIMESTAMP_STOP,
 271                                                cpu_to_be64(tstamp->stop))))
 272                 goto nla_put_failure;
 273         nla_nest_end(skb, nest_count);
 274
 275         return 0;
 276
 277 nla_put_failure:
 278         return -1;
 279 }
 280
 281 #ifdef CONFIG_NF_CONNTRACK_MARK
 282 static inline int
 283 ctnetlink_dump_mark(struct sk_buff *skb, const struct nf_conn *ct)
 284 {
 285         if (nla_put_be32(skb, CTA_MARK, htonl(ct->mark)))
 286                 goto nla_put_failure;
 287         return 0;
 288
 289 nla_put_failure:
 290         return -1;
 291 }
 292 #else
 293 #define ctnetlink_dump_mark(a, b) (0)
 294 #endif
 295
 296 #ifdef CONFIG_NF_CONNTRACK_SECMARK
 297 static inline int
 298 ctnetlink_dump_secctx(struct sk_buff *skb, const struct nf_conn *ct)
 299 {
 300         struct nlattr *nest_secctx;
 301         int len, ret;
 302         char *secctx;
 303
 304         ret = security_secid_to_secctx(ct->secmark, &secctx, &len);
 305         if (ret)
 306                 return 0;
 307
 308         ret = -1;
 309         nest_secctx = nla_nest_start(skb, CTA_SECCTX | NLA_F_NESTED);
 310         if (!nest_secctx)
 311                 goto nla_put_failure;
 312
 313         if (nla_put_string(skb, CTA_SECCTX_NAME, secctx))
 314                 goto nla_put_failure;
 315         nla_nest_end(skb, nest_secctx);
 316
 317         ret = 0;
 318 nla_put_failure:
 319         security_release_secctx(secctx, len);
 320         return ret;
 321 }
 322 #else
 323 #define ctnetlink_dump_secctx(a, b) (0)
 324 #endif
 325
 326 #define master_tuple(ct) &(ct->master->tuplehash[IP_CT_DIR_ORIGINAL].tuple)
 327
 328 static inline int
 329 ctnetlink_dump_master(struct sk_buff *skb, const struct nf_conn *ct)
 330 {
 331         struct nlattr *nest_parms;
 332
 333         if (!(ct->status & IPS_EXPECTED))
 334                 return 0;
 335
 336         nest_parms = nla_nest_start(skb, CTA_TUPLE_MASTER | NLA_F_NESTED);
 337         if (!nest_parms)
 338                 goto nla_put_failure;
 339         if (ctnetlink_dump_tuples(skb, master_tuple(ct)) < 0)
 340                 goto nla_put_failure;
 341         nla_nest_end(skb, nest_parms);
 342
 343         return 0;
 344
 345 nla_put_failure:
 346         return -1;
 347 }
 348
 349 #ifdef CONFIG_NF_NAT_NEEDED
 350 static int
 351 dump_nat_seq_adj(struct sk_buff *skb, const struct nf_nat_seq *natseq, int type)
 352 {
 353         struct nlattr *nest_parms;
 354
 355         nest_parms = nla_nest_start(skb, type | NLA_F_NESTED);
 356         if (!nest_parms)
 357                 goto nla_put_failure;
 358
 359         if (nla_put_be32(skb, CTA_NAT_SEQ_CORRECTION_POS,
 360                          htonl(natseq->correction_pos)) ||
 361             nla_put_be32(skb, CTA_NAT_SEQ_OFFSET_BEFORE,
 362                          htonl(natseq->offset_before)) ||
 363             nla_put_be32(skb, CTA_NAT_SEQ_OFFSET_AFTER,
 364                          htonl(natseq->offset_after)))
 365                 goto nla_put_failure;
 366
 367         nla_nest_end(skb, nest_parms);
 368
 369         return 0;
 370
 371 nla_put_failure:
 372         return -1;
 373 }
 374
 375 static inline int
 376 ctnetlink_dump_nat_seq_adj(struct sk_buff *skb, const struct nf_conn *ct)
 377 {
 378         struct nf_nat_seq *natseq;
 379         struct nf_conn_nat *nat = nfct_nat(ct);
 380
 381         if (!(ct->status & IPS_SEQ_ADJUST) || !nat)
 382                 return 0;
 383
 384         natseq = &nat->seq[IP_CT_DIR_ORIGINAL];
 385         if (dump_nat_seq_adj(skb, natseq, CTA_NAT_SEQ_ADJ_ORIG) == -1)
 386                 return -1;
 387
 388         natseq = &nat->seq[IP_CT_DIR_REPLY];
 389         if (dump_nat_seq_adj(skb, natseq, CTA_NAT_SEQ_ADJ_REPLY) == -1)
 390                 return -1;
 391
 392         return 0;
 393 }
 394 #else
 395 #define ctnetlink_dump_nat_seq_adj(a, b) (0)
 396 #endif
 397
 398 static inline int
 399 ctnetlink_dump_id(struct sk_buff *skb, const struct nf_conn *ct)
 400 {
 401         if (nla_put_be32(skb, CTA_ID, htonl((unsigned long)ct)))
 402                 goto nla_put_failure;
 403         return 0;
 404
 405 nla_put_failure:
 406         return -1;
 407 }
 408
 409 static inline int
 410 ctnetlink_dump_use(struct sk_buff *skb, const struct nf_conn *ct)
 411 {
 412         if (nla_put_be32(skb, CTA_USE, htonl(atomic_read(&ct->ct_general.use))))
 413                 goto nla_put_failure;
 414         return 0;
 415
 416 nla_put_failure:
 417         return -1;
 418 }
 419
 420 static int
 421 ctnetlink_fill_info(struct sk_buff *skb, u32 portid, u32 seq, u32 type,
 422                     struct nf_conn *ct)
 423 {
 424         struct nlmsghdr *nlh;
 425         struct nfgenmsg *nfmsg;
 426         struct nlattr *nest_parms;
 427         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
 428
 429         event = (NFNL_SUBSYS_CTNETLINK << 8 | IPCTNL_MSG_CT_NEW);
 430         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
 431         if (nlh == NULL)
 432                 goto nlmsg_failure;
 433
 434         nfmsg = nlmsg_data(nlh);
 435         nfmsg->nfgen_family = nf_ct_l3num(ct);
 436         nfmsg->version      = NFNETLINK_V0;
 437         nfmsg->res_id       = 0;
 438
 439         nest_parms = nla_nest_start(skb, CTA_TUPLE_ORIG | NLA_F_NESTED);
 440         if (!nest_parms)
 441                 goto nla_put_failure;
 442         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL)) < 0)
 443                 goto nla_put_failure;
 444         nla_nest_end(skb, nest_parms);
 445
 446         nest_parms = nla_nest_start(skb, CTA_TUPLE_REPLY | NLA_F_NESTED);
 447         if (!nest_parms)
 448                 goto nla_put_failure;
 449         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_REPLY)) < 0)
 450                 goto nla_put_failure;
 451         nla_nest_end(skb, nest_parms);
 452
 453         if (nf_ct_zone(ct) &&
 454             nla_put_be16(skb, CTA_ZONE, htons(nf_ct_zone(ct))))
 455                 goto nla_put_failure;
 456
 457         if (ctnetlink_dump_status(skb, ct) < 0 ||
 458             ctnetlink_dump_timeout(skb, ct) < 0 ||
 459             ctnetlink_dump_counters(skb, ct, IP_CT_DIR_ORIGINAL, type) < 0 ||
 460             ctnetlink_dump_counters(skb, ct, IP_CT_DIR_REPLY, type) < 0 ||
 461             ctnetlink_dump_timestamp(skb, ct) < 0 ||
 462             ctnetlink_dump_protoinfo(skb, ct) < 0 ||
 463             ctnetlink_dump_helpinfo(skb, ct) < 0 ||
 464             ctnetlink_dump_mark(skb, ct) < 0 ||
 465             ctnetlink_dump_secctx(skb, ct) < 0 ||
 466             ctnetlink_dump_id(skb, ct) < 0 ||
 467             ctnetlink_dump_use(skb, ct) < 0 ||
 468             ctnetlink_dump_master(skb, ct) < 0 ||
 469             ctnetlink_dump_nat_seq_adj(skb, ct) < 0)
 470                 goto nla_put_failure;
 471
 472         nlmsg_end(skb, nlh);
 473         return skb->len;
 474
 475 nlmsg_failure:
 476 nla_put_failure:
 477         nlmsg_cancel(skb, nlh);
 478         return -1;
 479 }
 480
 481 static inline size_t
 482 ctnetlink_proto_size(const struct nf_conn *ct)
 483 {
 484         struct nf_conntrack_l3proto *l3proto;
 485         struct nf_conntrack_l4proto *l4proto;
 486         size_t len = 0;
 487
 488         rcu_read_lock();
 489         l3proto = __nf_ct_l3proto_find(nf_ct_l3num(ct));
 490         len += l3proto->nla_size;
 491
 492         l4proto = __nf_ct_l4proto_find(nf_ct_l3num(ct), nf_ct_protonum(ct));
 493         len += l4proto->nla_size;
 494         rcu_read_unlock();
 495
 496         return len;
 497 }
 498
 499 static inline size_t
 500 ctnetlink_counters_size(const struct nf_conn *ct)
 501 {
 502         if (!nf_ct_ext_exist(ct, NF_CT_EXT_ACCT))
 503                 return 0;
 504         return 2 * nla_total_size(0) /* CTA_COUNTERS_ORIG|REPL */
 505                + 2 * nla_total_size(sizeof(uint64_t)) /* CTA_COUNTERS_PACKETS */
 506                + 2 * nla_total_size(sizeof(uint64_t)) /* CTA_COUNTERS_BYTES */
 507                ;
 508 }
 509
 510 static inline int
 511 ctnetlink_secctx_size(const struct nf_conn *ct)
 512 {
 513 #ifdef CONFIG_NF_CONNTRACK_SECMARK
 514         int len, ret;
 515
 516         ret = security_secid_to_secctx(ct->secmark, NULL, &len);
 517         if (ret)
 518                 return 0;
 519
 520         return nla_total_size(0) /* CTA_SECCTX */
 521                + nla_total_size(sizeof(char) * len); /* CTA_SECCTX_NAME */
 522 #else
 523         return 0;
 524 #endif
 525 }
 526
 527 static inline size_t
 528 ctnetlink_timestamp_size(const struct nf_conn *ct)
 529 {
 530 #ifdef CONFIG_NF_CONNTRACK_TIMESTAMP
 531         if (!nf_ct_ext_exist(ct, NF_CT_EXT_TSTAMP))
 532                 return 0;
 533         return nla_total_size(0) + 2 * nla_total_size(sizeof(uint64_t));
 534 #else
 535         return 0;
 536 #endif
 537 }
 538
 539 static inline size_t
 540 ctnetlink_nlmsg_size(const struct nf_conn *ct)
 541 {
 542         return NLMSG_ALIGN(sizeof(struct nfgenmsg))
 543                + 3 * nla_total_size(0) /* CTA_TUPLE_ORIG|REPL|MASTER */
 544                + 3 * nla_total_size(0) /* CTA_TUPLE_IP */
 545                + 3 * nla_total_size(0) /* CTA_TUPLE_PROTO */
 546                + 3 * nla_total_size(sizeof(u_int8_t)) /* CTA_PROTO_NUM */
 547                + nla_total_size(sizeof(u_int32_t)) /* CTA_ID */
 548                + nla_total_size(sizeof(u_int32_t)) /* CTA_STATUS */
 549                + ctnetlink_counters_size(ct)
 550                + ctnetlink_timestamp_size(ct)
 551                + nla_total_size(sizeof(u_int32_t)) /* CTA_TIMEOUT */
 552                + nla_total_size(0) /* CTA_PROTOINFO */
 553                + nla_total_size(0) /* CTA_HELP */
 554                + nla_total_size(NF_CT_HELPER_NAME_LEN) /* CTA_HELP_NAME */
 555                + ctnetlink_secctx_size(ct)
 556 #ifdef CONFIG_NF_NAT_NEEDED
 557                + 2 * nla_total_size(0) /* CTA_NAT_SEQ_ADJ_ORIG|REPL */
 558                + 6 * nla_total_size(sizeof(u_int32_t)) /* CTA_NAT_SEQ_OFFSET */
 559 #endif
 560 #ifdef CONFIG_NF_CONNTRACK_MARK
 561                + nla_total_size(sizeof(u_int32_t)) /* CTA_MARK */
 562 #endif
 563                + ctnetlink_proto_size(ct)
 564                ;
 565 }
 566
 567 #ifdef CONFIG_NF_CONNTRACK_EVENTS
 568 static int
 569 ctnetlink_conntrack_event(unsigned int events, struct nf_ct_event *item)
 570 {
 571         struct net *net;
 572         struct nlmsghdr *nlh;
 573         struct nfgenmsg *nfmsg;
 574         struct nlattr *nest_parms;
 575         struct nf_conn *ct = item->ct;
 576         struct sk_buff *skb;
 577         unsigned int type;
 578         unsigned int flags = 0, group;
 579         int err;
 580
 581         /* ignore our fake conntrack entry */
 582         if (nf_ct_is_untracked(ct))
 583                 return 0;
 584
 585         if (events & (1 << IPCT_DESTROY)) {
 586                 type = IPCTNL_MSG_CT_DELETE;
 587                 group = NFNLGRP_CONNTRACK_DESTROY;
 588         } else  if (events & ((1 << IPCT_NEW) | (1 << IPCT_RELATED))) {
 589                 type = IPCTNL_MSG_CT_NEW;
 590                 flags = NLM_F_CREATE|NLM_F_EXCL;
 591                 group = NFNLGRP_CONNTRACK_NEW;
 592         } else  if (events) {
 593                 type = IPCTNL_MSG_CT_NEW;
 594                 group = NFNLGRP_CONNTRACK_UPDATE;
 595         } else
 596                 return 0;
 597
 598         net = nf_ct_net(ct);
 599         if (!item->report && !nfnetlink_has_listeners(net, group))
 600                 return 0;
 601
 602         skb = nlmsg_new(ctnetlink_nlmsg_size(ct), GFP_ATOMIC);
 603         if (skb == NULL)
 604                 goto errout;
 605
 606         type |= NFNL_SUBSYS_CTNETLINK << 8;
 607         nlh = nlmsg_put(skb, item->portid, 0, type, sizeof(*nfmsg), flags);
 608         if (nlh == NULL)
 609                 goto nlmsg_failure;
 610
 611         nfmsg = nlmsg_data(nlh);
 612         nfmsg->nfgen_family = nf_ct_l3num(ct);
 613         nfmsg->version  = NFNETLINK_V0;
 614         nfmsg->res_id   = 0;
 615
 616         rcu_read_lock();
 617         nest_parms = nla_nest_start(skb, CTA_TUPLE_ORIG | NLA_F_NESTED);
 618         if (!nest_parms)
 619                 goto nla_put_failure;
 620         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL)) < 0)
 621                 goto nla_put_failure;
 622         nla_nest_end(skb, nest_parms);
 623
 624         nest_parms = nla_nest_start(skb, CTA_TUPLE_REPLY | NLA_F_NESTED);
 625         if (!nest_parms)
 626                 goto nla_put_failure;
 627         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_REPLY)) < 0)
 628                 goto nla_put_failure;
 629         nla_nest_end(skb, nest_parms);
 630
 631         if (nf_ct_zone(ct) &&
 632             nla_put_be16(skb, CTA_ZONE, htons(nf_ct_zone(ct))))
 633                 goto nla_put_failure;
 634
 635         if (ctnetlink_dump_id(skb, ct) < 0)
 636                 goto nla_put_failure;
 637
 638         if (ctnetlink_dump_status(skb, ct) < 0)
 639                 goto nla_put_failure;
 640
 641         if (events & (1 << IPCT_DESTROY)) {
 642                 if (ctnetlink_dump_counters(skb, ct,
 643                                             IP_CT_DIR_ORIGINAL, type) < 0 ||
 644                     ctnetlink_dump_counters(skb, ct,
 645                                             IP_CT_DIR_REPLY, type) < 0 ||
 646                     ctnetlink_dump_timestamp(skb, ct) < 0)
 647                         goto nla_put_failure;
 648         } else {
 649                 if (ctnetlink_dump_timeout(skb, ct) < 0)
 650                         goto nla_put_failure;
 651
 652                 if (events & (1 << IPCT_PROTOINFO)
 653                     && ctnetlink_dump_protoinfo(skb, ct) < 0)
 654                         goto nla_put_failure;
 655
 656                 if ((events & (1 << IPCT_HELPER) || nfct_help(ct))
 657                     && ctnetlink_dump_helpinfo(skb, ct) < 0)
 658                         goto nla_put_failure;
 659
 660 #ifdef CONFIG_NF_CONNTRACK_SECMARK
 661                 if ((events & (1 << IPCT_SECMARK) || ct->secmark)
 662                     && ctnetlink_dump_secctx(skb, ct) < 0)
 663                         goto nla_put_failure;
 664 #endif
 665
 666                 if (events & (1 << IPCT_RELATED) &&
 667                     ctnetlink_dump_master(skb, ct) < 0)
 668                         goto nla_put_failure;
 669
 670                 if (events & (1 << IPCT_NATSEQADJ) &&
 671                     ctnetlink_dump_nat_seq_adj(skb, ct) < 0)
 672                         goto nla_put_failure;
 673         }
 674
 675 #ifdef CONFIG_NF_CONNTRACK_MARK
 676         if ((events & (1 << IPCT_MARK) || ct->mark)
 677             && ctnetlink_dump_mark(skb, ct) < 0)
 678                 goto nla_put_failure;
 679 #endif
 680         rcu_read_unlock();
 681
 682         nlmsg_end(skb, nlh);
 683         err = nfnetlink_send(skb, net, item->portid, group, item->report,
 684                              GFP_ATOMIC);
 685         if (err == -ENOBUFS || err == -EAGAIN)
 686                 return -ENOBUFS;
 687
 688         return 0;
 689
 690 nla_put_failure:
 691         rcu_read_unlock();
 692         nlmsg_cancel(skb, nlh);
 693 nlmsg_failure:
 694         kfree_skb(skb);
 695 errout:
 696         if (nfnetlink_set_err(net, 0, group, -ENOBUFS) > 0)
 697                 return -ENOBUFS;
 698
 699         return 0;
 700 }
 701 #endif /* CONFIG_NF_CONNTRACK_EVENTS */
 702
 703 static int ctnetlink_done(struct netlink_callback *cb)
 704 {
 705         if (cb->args[1])
 706                 nf_ct_put((struct nf_conn *)cb->args[1]);
 707         if (cb->data)
 708                 kfree(cb->data);
 709         return 0;
 710 }
 711
 712 struct ctnetlink_dump_filter {
 713         struct {
 714                 u_int32_t val;
 715                 u_int32_t mask;
 716         } mark;
 717 };
 718
 719 static int
 720 ctnetlink_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
 721 {
 722         struct net *net = sock_net(skb->sk);
 723         struct nf_conn *ct, *last;
 724         struct nf_conntrack_tuple_hash *h;
 725         struct hlist_nulls_node *n;
 726         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
 727         u_int8_t l3proto = nfmsg->nfgen_family;
 728         int res;
 729 #ifdef CONFIG_NF_CONNTRACK_MARK
 730         const struct ctnetlink_dump_filter *filter = cb->data;
 731 #endif
 732
 733         spin_lock_bh(&nf_conntrack_lock);
 734         last = (struct nf_conn *)cb->args[1];
 735         for (; cb->args[0] < net->ct.htable_size; cb->args[0]++) {
 736 restart:
 737                 hlist_nulls_for_each_entry(h, n, &net->ct.hash[cb->args[0]],
 738                                          hnnode) {
 739                         if (NF_CT_DIRECTION(h) != IP_CT_DIR_ORIGINAL)
 740                                 continue;
 741                         ct = nf_ct_tuplehash_to_ctrack(h);
 742                         /* Dump entries of a given L3 protocol number.
 743                          * If it is not specified, ie. l3proto == 0,
 744                          * then dump everything. */
 745                         if (l3proto && nf_ct_l3num(ct) != l3proto)
 746                                 continue;
 747                         if (cb->args[1]) {
 748                                 if (ct != last)
 749                                         continue;
 750                                 cb->args[1] = 0;
 751                         }
 752 #ifdef CONFIG_NF_CONNTRACK_MARK
 753                         if (filter && !((ct->mark & filter->mark.mask) ==
 754                                         filter->mark.val)) {
 755                                 continue;
 756                         }
 757 #endif
 758                         rcu_read_lock();
 759                         res =
 760                         ctnetlink_fill_info(skb, NETLINK_CB(cb->skb).portid,
 761                                             cb->nlh->nlmsg_seq,
 762                                             NFNL_MSG_TYPE(cb->nlh->nlmsg_type),
 763                                             ct);
 764                         rcu_read_unlock();
 765                         if (res < 0) {
 766                                 nf_conntrack_get(&ct->ct_general);
 767                                 cb->args[1] = (unsigned long)ct;
 768                                 goto out;
 769                         }
 770                 }
 771                 if (cb->args[1]) {
 772                         cb->args[1] = 0;
 773                         goto restart;
 774                 }
 775         }
 776 out:
 777         spin_unlock_bh(&nf_conntrack_lock);
 778         if (last)
 779                 nf_ct_put(last);
 780
 781         return skb->len;
 782 }
 783
 784 static inline int
 785 ctnetlink_parse_tuple_ip(struct nlattr *attr, struct nf_conntrack_tuple *tuple)
 786 {
 787         struct nlattr *tb[CTA_IP_MAX+1];
 788         struct nf_conntrack_l3proto *l3proto;
 789         int ret = 0;
 790
 791         nla_parse_nested(tb, CTA_IP_MAX, attr, NULL);
 792
 793         rcu_read_lock();
 794         l3proto = __nf_ct_l3proto_find(tuple->src.l3num);
 795
 796         if (likely(l3proto->nlattr_to_tuple)) {
 797                 ret = nla_validate_nested(attr, CTA_IP_MAX,
 798                                           l3proto->nla_policy);
 799                 if (ret == 0)
 800                         ret = l3proto->nlattr_to_tuple(tb, tuple);
 801         }
 802
 803         rcu_read_unlock();
 804
 805         return ret;
 806 }
 807
 808 static const struct nla_policy proto_nla_policy[CTA_PROTO_MAX+1] = {
 809         [CTA_PROTO_NUM] = { .type = NLA_U8 },
 810 };
 811
 812 static inline int
 813 ctnetlink_parse_tuple_proto(struct nlattr *attr,
 814                             struct nf_conntrack_tuple *tuple)
 815 {
 816         struct nlattr *tb[CTA_PROTO_MAX+1];
 817         struct nf_conntrack_l4proto *l4proto;
 818         int ret = 0;
 819
 820         ret = nla_parse_nested(tb, CTA_PROTO_MAX, attr, proto_nla_policy);
 821         if (ret < 0)
 822                 return ret;
 823
 824         if (!tb[CTA_PROTO_NUM])
 825                 return -EINVAL;
 826         tuple->dst.protonum = nla_get_u8(tb[CTA_PROTO_NUM]);
 827
 828         rcu_read_lock();
 829         l4proto = __nf_ct_l4proto_find(tuple->src.l3num, tuple->dst.protonum);
 830
 831         if (likely(l4proto->nlattr_to_tuple)) {
 832                 ret = nla_validate_nested(attr, CTA_PROTO_MAX,
 833                                           l4proto->nla_policy);
 834                 if (ret == 0)
 835                         ret = l4proto->nlattr_to_tuple(tb, tuple);
 836         }
 837
 838         rcu_read_unlock();
 839
 840         return ret;
 841 }
 842
 843 static const struct nla_policy tuple_nla_policy[CTA_TUPLE_MAX+1] = {
 844         [CTA_TUPLE_IP]          = { .type = NLA_NESTED },
 845         [CTA_TUPLE_PROTO]       = { .type = NLA_NESTED },
 846 };
 847
 848 static int
 849 ctnetlink_parse_tuple(const struct nlattr * const cda[],
 850                       struct nf_conntrack_tuple *tuple,
 851                       enum ctattr_type type, u_int8_t l3num)
 852 {
 853         struct nlattr *tb[CTA_TUPLE_MAX+1];
 854         int err;
 855
 856         memset(tuple, 0, sizeof(*tuple));
 857
 858         nla_parse_nested(tb, CTA_TUPLE_MAX, cda[type], tuple_nla_policy);
 859
 860         if (!tb[CTA_TUPLE_IP])
 861                 return -EINVAL;
 862
 863         tuple->src.l3num = l3num;
 864
 865         err = ctnetlink_parse_tuple_ip(tb[CTA_TUPLE_IP], tuple);
 866         if (err < 0)
 867                 return err;
 868
 869         if (!tb[CTA_TUPLE_PROTO])
 870                 return -EINVAL;
 871
 872         err = ctnetlink_parse_tuple_proto(tb[CTA_TUPLE_PROTO], tuple);
 873         if (err < 0)
 874                 return err;
 875
 876         /* orig and expect tuples get DIR_ORIGINAL */
 877         if (type == CTA_TUPLE_REPLY)
 878                 tuple->dst.dir = IP_CT_DIR_REPLY;
 879         else
 880                 tuple->dst.dir = IP_CT_DIR_ORIGINAL;
 881
 882         return 0;
 883 }
 884
 885 static int
 886 ctnetlink_parse_zone(const struct nlattr *attr, u16 *zone)
 887 {
 888         if (attr)
 889 #ifdef CONFIG_NF_CONNTRACK_ZONES
 890                 *zone = ntohs(nla_get_be16(attr));
 891 #else
 892                 return -EOPNOTSUPP;
 893 #endif
 894         else
 895                 *zone = 0;
 896
 897         return 0;
 898 }
 899
 900 static const struct nla_policy help_nla_policy[CTA_HELP_MAX+1] = {
 901         [CTA_HELP_NAME]         = { .type = NLA_NUL_STRING },
 902 };
 903
 904 static inline int
 905 ctnetlink_parse_help(const struct nlattr *attr, char **helper_name,
 906                      struct nlattr **helpinfo)
 907 {
 908         struct nlattr *tb[CTA_HELP_MAX+1];
 909
 910         nla_parse_nested(tb, CTA_HELP_MAX, attr, help_nla_policy);
 911
 912         if (!tb[CTA_HELP_NAME])
 913                 return -EINVAL;
 914
 915         *helper_name = nla_data(tb[CTA_HELP_NAME]);
 916
 917         if (tb[CTA_HELP_INFO])
 918                 *helpinfo = tb[CTA_HELP_INFO];
 919
 920         return 0;
 921 }
 922
 923 static const struct nla_policy ct_nla_policy[CTA_MAX+1] = {
 924         [CTA_TUPLE_ORIG]        = { .type = NLA_NESTED },
 925         [CTA_TUPLE_REPLY]       = { .type = NLA_NESTED },
 926         [CTA_STATUS]            = { .type = NLA_U32 },
 927         [CTA_PROTOINFO]         = { .type = NLA_NESTED },
 928         [CTA_HELP]              = { .type = NLA_NESTED },
 929         [CTA_NAT_SRC]           = { .type = NLA_NESTED },
 930         [CTA_TIMEOUT]           = { .type = NLA_U32 },
 931         [CTA_MARK]              = { .type = NLA_U32 },
 932         [CTA_ID]                = { .type = NLA_U32 },
 933         [CTA_NAT_DST]           = { .type = NLA_NESTED },
 934         [CTA_TUPLE_MASTER]      = { .type = NLA_NESTED },
 935         [CTA_ZONE]              = { .type = NLA_U16 },
 936         [CTA_MARK_MASK]         = { .type = NLA_U32 },
 937 };
 938
 939 static int
 940 ctnetlink_del_conntrack(struct sock *ctnl, struct sk_buff *skb,
 941                         const struct nlmsghdr *nlh,
 942                         const struct nlattr * const cda[])
 943 {
 944         struct net *net = sock_net(ctnl);
 945         struct nf_conntrack_tuple_hash *h;
 946         struct nf_conntrack_tuple tuple;
 947         struct nf_conn *ct;
 948         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
 949         u_int8_t u3 = nfmsg->nfgen_family;
 950         u16 zone;
 951         int err;
 952
 953         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
 954         if (err < 0)
 955                 return err;
 956
 957         if (cda[CTA_TUPLE_ORIG])
 958                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_ORIG, u3);
 959         else if (cda[CTA_TUPLE_REPLY])
 960                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_REPLY, u3);
 961         else {
 962                 /* Flush the whole table */
 963                 nf_conntrack_flush_report(net,
 964                                          NETLINK_CB(skb).portid,
 965                                          nlmsg_report(nlh));
 966                 return 0;
 967         }
 968
 969         if (err < 0)
 970                 return err;
 971
 972         h = nf_conntrack_find_get(net, zone, &tuple);
 973         if (!h)
 974                 return -ENOENT;
 975
 976         ct = nf_ct_tuplehash_to_ctrack(h);
 977
 978         if (cda[CTA_ID]) {
 979                 u_int32_t id = ntohl(nla_get_be32(cda[CTA_ID]));
 980                 if (id != (u32)(unsigned long)ct) {
 981                         nf_ct_put(ct);
 982                         return -ENOENT;
 983                 }
 984         }
 985
 986         if (del_timer(&ct->timeout)) {
 987                 if (nf_conntrack_event_report(IPCT_DESTROY, ct,
 988                                               NETLINK_CB(skb).portid,
 989                                               nlmsg_report(nlh)) < 0) {
 990                         nf_ct_delete_from_lists(ct);
 991                         /* we failed to report the event, try later */
 992                         nf_ct_dying_timeout(ct);
 993                         nf_ct_put(ct);
 994                         return 0;
 995                 }
 996                 /* death_by_timeout would report the event again */
 997                 set_bit(IPS_DYING_BIT, &ct->status);
 998                 nf_ct_delete_from_lists(ct);
 999                 nf_ct_put(ct);
1000         }
1001         nf_ct_put(ct);
1002
1003         return 0;
1004 }
1005
1006 static int
1007 ctnetlink_get_conntrack(struct sock *ctnl, struct sk_buff *skb,
1008                         const struct nlmsghdr *nlh,
1009                         const struct nlattr * const cda[])
1010 {
1011         struct net *net = sock_net(ctnl);
1012         struct nf_conntrack_tuple_hash *h;
1013         struct nf_conntrack_tuple tuple;
1014         struct nf_conn *ct;
1015         struct sk_buff *skb2 = NULL;
1016         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
1017         u_int8_t u3 = nfmsg->nfgen_family;
1018         u16 zone;
1019         int err;
1020
1021         if (nlh->nlmsg_flags & NLM_F_DUMP) {
1022                 struct netlink_dump_control c = {
1023                         .dump = ctnetlink_dump_table,
1024                         .done = ctnetlink_done,
1025                 };
1026 #ifdef CONFIG_NF_CONNTRACK_MARK
1027                 if (cda[CTA_MARK] && cda[CTA_MARK_MASK]) {
1028                         struct ctnetlink_dump_filter *filter;
1029
1030                         filter = kzalloc(sizeof(struct ctnetlink_dump_filter),
1031                                          GFP_ATOMIC);
1032                         if (filter == NULL)
1033                                 return -ENOMEM;
1034
1035                         filter->mark.val = ntohl(nla_get_be32(cda[CTA_MARK]));
1036                         filter->mark.mask =
1037                                 ntohl(nla_get_be32(cda[CTA_MARK_MASK]));
1038                         c.data = filter;
1039                 }
1040 #endif
1041                 return netlink_dump_start(ctnl, skb, nlh, &c);
1042         }
1043
1044         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
1045         if (err < 0)
1046                 return err;
1047
1048         if (cda[CTA_TUPLE_ORIG])
1049                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_ORIG, u3);
1050         else if (cda[CTA_TUPLE_REPLY])
1051                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_TUPLE_REPLY, u3);
1052         else
1053                 return -EINVAL;
1054
1055         if (err < 0)
1056                 return err;
1057
1058         h = nf_conntrack_find_get(net, zone, &tuple);
1059         if (!h)
1060                 return -ENOENT;
1061
1062         ct = nf_ct_tuplehash_to_ctrack(h);
1063
1064         err = -ENOMEM;
1065         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
1066         if (skb2 == NULL) {
1067                 nf_ct_put(ct);
1068                 return -ENOMEM;
1069         }
1070
1071         rcu_read_lock();
1072         err = ctnetlink_fill_info(skb2, NETLINK_CB(skb).portid, nlh->nlmsg_seq,
1073                                   NFNL_MSG_TYPE(nlh->nlmsg_type), ct);
1074         rcu_read_unlock();
1075         nf_ct_put(ct);
1076         if (err <= 0)
1077                 goto free;
1078
1079         err = netlink_unicast(ctnl, skb2, NETLINK_CB(skb).portid, MSG_DONTWAIT);
1080         if (err < 0)
1081                 goto out;
1082
1083         return 0;
1084
1085 free:
1086         kfree_skb(skb2);
1087 out:
1088         /* this avoids a loop in nfnetlink. */
1089         return err == -EAGAIN ? -ENOBUFS : err;
1090 }
1091
1092 static int ctnetlink_done_list(struct netlink_callback *cb)
1093 {
1094         if (cb->args[1])
1095                 nf_ct_put((struct nf_conn *)cb->args[1]);
1096         return 0;
1097 }
1098
1099 static int
1100 ctnetlink_dump_list(struct sk_buff *skb, struct netlink_callback *cb,
1101                     struct hlist_nulls_head *list)
1102 {
1103         struct nf_conn *ct, *last;
1104         struct nf_conntrack_tuple_hash *h;
1105         struct hlist_nulls_node *n;
1106         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
1107         u_int8_t l3proto = nfmsg->nfgen_family;
1108         int res;
1109
1110         if (cb->args[2])
1111                 return 0;
1112
1113         spin_lock_bh(&nf_conntrack_lock);
1114         last = (struct nf_conn *)cb->args[1];
1115 restart:
1116         hlist_nulls_for_each_entry(h, n, list, hnnode) {
1117                 ct = nf_ct_tuplehash_to_ctrack(h);
1118                 if (l3proto && nf_ct_l3num(ct) != l3proto)
1119                         continue;
1120                 if (cb->args[1]) {
1121                         if (ct != last)
1122                                 continue;
1123                         cb->args[1] = 0;
1124                 }
1125                 rcu_read_lock();
1126                 res = ctnetlink_fill_info(skb, NETLINK_CB(cb->skb).portid,
1127                                           cb->nlh->nlmsg_seq,
1128                                           NFNL_MSG_TYPE(cb->nlh->nlmsg_type),
1129                                           ct);
1130                 rcu_read_unlock();
1131                 if (res < 0) {
1132                         nf_conntrack_get(&ct->ct_general);
1133                         cb->args[1] = (unsigned long)ct;
1134                         goto out;
1135                 }
1136         }
1137         if (cb->args[1]) {
1138                 cb->args[1] = 0;
1139                 goto restart;
1140         } else
1141                 cb->args[2] = 1;
1142 out:
1143         spin_unlock_bh(&nf_conntrack_lock);
1144         if (last)
1145                 nf_ct_put(last);
1146
1147         return skb->len;
1148 }
1149
1150 static int
1151 ctnetlink_dump_dying(struct sk_buff *skb, struct netlink_callback *cb)
1152 {
1153         struct net *net = sock_net(skb->sk);
1154
1155         return ctnetlink_dump_list(skb, cb, &net->ct.dying);
1156 }
1157
1158 static int
1159 ctnetlink_get_ct_dying(struct sock *ctnl, struct sk_buff *skb,
1160                        const struct nlmsghdr *nlh,
1161                        const struct nlattr * const cda[])
1162 {
1163         if (nlh->nlmsg_flags & NLM_F_DUMP) {
1164                 struct netlink_dump_control c = {
1165                         .dump = ctnetlink_dump_dying,
1166                         .done = ctnetlink_done_list,
1167                 };
1168                 return netlink_dump_start(ctnl, skb, nlh, &c);
1169         }
1170
1171         return -EOPNOTSUPP;
1172 }
1173
1174 static int
1175 ctnetlink_dump_unconfirmed(struct sk_buff *skb, struct netlink_callback *cb)
1176 {
1177         struct net *net = sock_net(skb->sk);
1178
1179         return ctnetlink_dump_list(skb, cb, &net->ct.unconfirmed);
1180 }
1181
1182 static int
1183 ctnetlink_get_ct_unconfirmed(struct sock *ctnl, struct sk_buff *skb,
1184                              const struct nlmsghdr *nlh,
1185                              const struct nlattr * const cda[])
1186 {
1187         if (nlh->nlmsg_flags & NLM_F_DUMP) {
1188                 struct netlink_dump_control c = {
1189                         .dump = ctnetlink_dump_unconfirmed,
1190                         .done = ctnetlink_done_list,
1191                 };
1192                 return netlink_dump_start(ctnl, skb, nlh, &c);
1193         }
1194
1195         return -EOPNOTSUPP;
1196 }
1197
1198 #ifdef CONFIG_NF_NAT_NEEDED
1199 static int
1200 ctnetlink_parse_nat_setup(struct nf_conn *ct,
1201                           enum nf_nat_manip_type manip,
1202                           const struct nlattr *attr)
1203 {
1204         typeof(nfnetlink_parse_nat_setup_hook) parse_nat_setup;
1205         int err;
1206
1207         parse_nat_setup = rcu_dereference(nfnetlink_parse_nat_setup_hook);
1208         if (!parse_nat_setup) {
1209 #ifdef CONFIG_MODULES
1210                 rcu_read_unlock();
1211                 nfnl_unlock();
1212                 if (request_module("nf-nat") < 0) {
1213                         nfnl_lock();
1214                         rcu_read_lock();
1215                         return -EOPNOTSUPP;
1216                 }
1217                 nfnl_lock();
1218                 rcu_read_lock();
1219                 if (nfnetlink_parse_nat_setup_hook)
1220                         return -EAGAIN;
1221 #endif
1222                 return -EOPNOTSUPP;
1223         }
1224
1225         err = parse_nat_setup(ct, manip, attr);
1226         if (err == -EAGAIN) {
1227 #ifdef CONFIG_MODULES
1228                 rcu_read_unlock();
1229                 nfnl_unlock();
1230                 if (request_module("nf-nat-%u", nf_ct_l3num(ct)) < 0) {
1231                         nfnl_lock();
1232                         rcu_read_lock();
1233                         return -EOPNOTSUPP;
1234                 }
1235                 nfnl_lock();
1236                 rcu_read_lock();
1237 #else
1238                 err = -EOPNOTSUPP;
1239 #endif
1240         }
1241         return err;
1242 }
1243 #endif
1244
1245 static int
1246 ctnetlink_change_status(struct nf_conn *ct, const struct nlattr * const cda[])
1247 {
1248         unsigned long d;
1249         unsigned int status = ntohl(nla_get_be32(cda[CTA_STATUS]));
1250         d = ct->status ^ status;
1251
1252         if (d & (IPS_EXPECTED|IPS_CONFIRMED|IPS_DYING))
1253                 /* unchangeable */
1254                 return -EBUSY;
1255
1256         if (d & IPS_SEEN_REPLY && !(status & IPS_SEEN_REPLY))
1257                 /* SEEN_REPLY bit can only be set */
1258                 return -EBUSY;
1259
1260         if (d & IPS_ASSURED && !(status & IPS_ASSURED))
1261                 /* ASSURED bit can only be set */
1262                 return -EBUSY;
1263
1264         /* Be careful here, modifying NAT bits can screw up things,
1265          * so don't let users modify them directly if they don't pass
1266          * nf_nat_range. */
1267         ct->status |= status & ~(IPS_NAT_DONE_MASK | IPS_NAT_MASK);
1268         return 0;
1269 }
1270
1271 static int
1272 ctnetlink_change_nat(struct nf_conn *ct, const struct nlattr * const cda[])
1273 {
1274 #ifdef CONFIG_NF_NAT_NEEDED
1275         int ret;
1276
1277         if (cda[CTA_NAT_DST]) {
1278                 ret = ctnetlink_parse_nat_setup(ct,
1279                                                 NF_NAT_MANIP_DST,
1280                                                 cda[CTA_NAT_DST]);
1281                 if (ret < 0)
1282                         return ret;
1283         }
1284         if (cda[CTA_NAT_SRC]) {
1285                 ret = ctnetlink_parse_nat_setup(ct,
1286                                                 NF_NAT_MANIP_SRC,
1287                                                 cda[CTA_NAT_SRC]);
1288                 if (ret < 0)
1289                         return ret;
1290         }
1291         return 0;
1292 #else
1293         return -EOPNOTSUPP;
1294 #endif
1295 }
1296
1297 static inline int
1298 ctnetlink_change_helper(struct nf_conn *ct, const struct nlattr * const cda[])
1299 {
1300         struct nf_conntrack_helper *helper;
1301         struct nf_conn_help *help = nfct_help(ct);
1302         char *helpname = NULL;
1303         struct nlattr *helpinfo = NULL;
1304         int err;
1305
1306         /* don't change helper of sibling connections */
1307         if (ct->master)
1308                 return -EBUSY;
1309
1310         err = ctnetlink_parse_help(cda[CTA_HELP], &helpname, &helpinfo);
1311         if (err < 0)
1312                 return err;
1313
1314         if (!strcmp(helpname, "")) {
1315                 if (help && help->helper) {
1316                         /* we had a helper before ... */
1317                         nf_ct_remove_expectations(ct);
1318                         RCU_INIT_POINTER(help->helper, NULL);
1319                 }
1320
1321                 return 0;
1322         }
1323
1324         helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
1325                                             nf_ct_protonum(ct));
1326         if (helper == NULL) {
1327 #ifdef CONFIG_MODULES
1328                 spin_unlock_bh(&nf_conntrack_lock);
1329
1330                 if (request_module("nfct-helper-%s", helpname) < 0) {
1331                         spin_lock_bh(&nf_conntrack_lock);
1332                         return -EOPNOTSUPP;
1333                 }
1334
1335                 spin_lock_bh(&nf_conntrack_lock);
1336                 helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
1337                                                     nf_ct_protonum(ct));
1338                 if (helper)
1339                         return -EAGAIN;
1340 #endif
1341                 return -EOPNOTSUPP;
1342         }
1343
1344         if (help) {
1345                 if (help->helper == helper) {
1346                         /* update private helper data if allowed. */
1347                         if (helper->from_nlattr)
1348                                 helper->from_nlattr(helpinfo, ct);
1349                         return 0;
1350                 } else
1351                         return -EBUSY;
1352         }
1353
1354         /* we cannot set a helper for an existing conntrack */
1355         return -EOPNOTSUPP;
1356 }
1357
1358 static inline int
1359 ctnetlink_change_timeout(struct nf_conn *ct, const struct nlattr * const cda[])
1360 {
1361         u_int32_t timeout = ntohl(nla_get_be32(cda[CTA_TIMEOUT]));
1362
1363         if (!del_timer(&ct->timeout))
1364                 return -ETIME;
1365
1366         ct->timeout.expires = jiffies + timeout * HZ;
1367         add_timer(&ct->timeout);
1368
1369         return 0;
1370 }
1371
1372 static const struct nla_policy protoinfo_policy[CTA_PROTOINFO_MAX+1] = {
1373         [CTA_PROTOINFO_TCP]     = { .type = NLA_NESTED },
1374         [CTA_PROTOINFO_DCCP]    = { .type = NLA_NESTED },
1375         [CTA_PROTOINFO_SCTP]    = { .type = NLA_NESTED },
1376 };
1377
1378 static inline int
1379 ctnetlink_change_protoinfo(struct nf_conn *ct, const struct nlattr * const cda[])
1380 {
1381         const struct nlattr *attr = cda[CTA_PROTOINFO];
1382         struct nlattr *tb[CTA_PROTOINFO_MAX+1];
1383         struct nf_conntrack_l4proto *l4proto;
1384         int err = 0;
1385
1386         nla_parse_nested(tb, CTA_PROTOINFO_MAX, attr, protoinfo_policy);
1387
1388         rcu_read_lock();
1389         l4proto = __nf_ct_l4proto_find(nf_ct_l3num(ct), nf_ct_protonum(ct));
1390         if (l4proto->from_nlattr)
1391                 err = l4proto->from_nlattr(tb, ct);
1392         rcu_read_unlock();
1393
1394         return err;
1395 }
1396
1397 #ifdef CONFIG_NF_NAT_NEEDED
1398 static const struct nla_policy nat_seq_policy[CTA_NAT_SEQ_MAX+1] = {
1399         [CTA_NAT_SEQ_CORRECTION_POS]    = { .type = NLA_U32 },
1400         [CTA_NAT_SEQ_OFFSET_BEFORE]     = { .type = NLA_U32 },
1401         [CTA_NAT_SEQ_OFFSET_AFTER]      = { .type = NLA_U32 },
1402 };
1403
1404 static inline int
1405 change_nat_seq_adj(struct nf_nat_seq *natseq, const struct nlattr * const attr)
1406 {
1407         struct nlattr *cda[CTA_NAT_SEQ_MAX+1];
1408
1409         nla_parse_nested(cda, CTA_NAT_SEQ_MAX, attr, nat_seq_policy);
1410
1411         if (!cda[CTA_NAT_SEQ_CORRECTION_POS])
1412                 return -EINVAL;
1413
1414         natseq->correction_pos =
1415                 ntohl(nla_get_be32(cda[CTA_NAT_SEQ_CORRECTION_POS]));
1416
1417         if (!cda[CTA_NAT_SEQ_OFFSET_BEFORE])
1418                 return -EINVAL;
1419
1420         natseq->offset_before =
1421                 ntohl(nla_get_be32(cda[CTA_NAT_SEQ_OFFSET_BEFORE]));
1422
1423         if (!cda[CTA_NAT_SEQ_OFFSET_AFTER])
1424                 return -EINVAL;
1425
1426         natseq->offset_after =
1427                 ntohl(nla_get_be32(cda[CTA_NAT_SEQ_OFFSET_AFTER]));
1428
1429         return 0;
1430 }
1431
1432 static int
1433 ctnetlink_change_nat_seq_adj(struct nf_conn *ct,
1434                              const struct nlattr * const cda[])
1435 {
1436         int ret = 0;
1437         struct nf_conn_nat *nat = nfct_nat(ct);
1438
1439         if (!nat)
1440                 return 0;
1441
1442         if (cda[CTA_NAT_SEQ_ADJ_ORIG]) {
1443                 ret = change_nat_seq_adj(&nat->seq[IP_CT_DIR_ORIGINAL],
1444                                          cda[CTA_NAT_SEQ_ADJ_ORIG]);
1445                 if (ret < 0)
1446                         return ret;
1447
1448                 ct->status |= IPS_SEQ_ADJUST;
1449         }
1450
1451         if (cda[CTA_NAT_SEQ_ADJ_REPLY]) {
1452                 ret = change_nat_seq_adj(&nat->seq[IP_CT_DIR_REPLY],
1453                                          cda[CTA_NAT_SEQ_ADJ_REPLY]);
1454                 if (ret < 0)
1455                         return ret;
1456
1457                 ct->status |= IPS_SEQ_ADJUST;
1458         }
1459
1460         return 0;
1461 }
1462 #endif
1463
1464 static int
1465 ctnetlink_change_conntrack(struct nf_conn *ct,
1466                            const struct nlattr * const cda[])
1467 {
1468         int err;
1469
1470         /* only allow NAT changes and master assignation for new conntracks */
1471         if (cda[CTA_NAT_SRC] || cda[CTA_NAT_DST] || cda[CTA_TUPLE_MASTER])
1472                 return -EOPNOTSUPP;
1473
1474         if (cda[CTA_HELP]) {
1475                 err = ctnetlink_change_helper(ct, cda);
1476                 if (err < 0)
1477                         return err;
1478         }
1479
1480         if (cda[CTA_TIMEOUT]) {
1481                 err = ctnetlink_change_timeout(ct, cda);
1482                 if (err < 0)
1483                         return err;
1484         }
1485
1486         if (cda[CTA_STATUS]) {
1487                 err = ctnetlink_change_status(ct, cda);
1488                 if (err < 0)
1489                         return err;
1490         }
1491
1492         if (cda[CTA_PROTOINFO]) {
1493                 err = ctnetlink_change_protoinfo(ct, cda);
1494                 if (err < 0)
1495                         return err;
1496         }
1497
1498 #if defined(CONFIG_NF_CONNTRACK_MARK)
1499         if (cda[CTA_MARK])
1500                 ct->mark = ntohl(nla_get_be32(cda[CTA_MARK]));
1501 #endif
1502
1503 #ifdef CONFIG_NF_NAT_NEEDED
1504         if (cda[CTA_NAT_SEQ_ADJ_ORIG] || cda[CTA_NAT_SEQ_ADJ_REPLY]) {
1505                 err = ctnetlink_change_nat_seq_adj(ct, cda);
1506                 if (err < 0)
1507                         return err;
1508         }
1509 #endif
1510
1511         return 0;
1512 }
1513
1514 static struct nf_conn *
1515 ctnetlink_create_conntrack(struct net *net, u16 zone,
1516                            const struct nlattr * const cda[],
1517                            struct nf_conntrack_tuple *otuple,
1518                            struct nf_conntrack_tuple *rtuple,
1519                            u8 u3)
1520 {
1521         struct nf_conn *ct;
1522         int err = -EINVAL;
1523         struct nf_conntrack_helper *helper;
1524         struct nf_conn_tstamp *tstamp;
1525
1526         ct = nf_conntrack_alloc(net, zone, otuple, rtuple, GFP_ATOMIC);
1527         if (IS_ERR(ct))
1528                 return ERR_PTR(-ENOMEM);
1529
1530         if (!cda[CTA_TIMEOUT])
1531                 goto err1;
1532         ct->timeout.expires = ntohl(nla_get_be32(cda[CTA_TIMEOUT]));
1533
1534         ct->timeout.expires = jiffies + ct->timeout.expires * HZ;
1535
1536         rcu_read_lock();
1537         if (cda[CTA_HELP]) {
1538                 char *helpname = NULL;
1539                 struct nlattr *helpinfo = NULL;
1540
1541                 err = ctnetlink_parse_help(cda[CTA_HELP], &helpname, &helpinfo);
1542                 if (err < 0)
1543                         goto err2;
1544
1545                 helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
1546                                                     nf_ct_protonum(ct));
1547                 if (helper == NULL) {
1548                         rcu_read_unlock();
1549 #ifdef CONFIG_MODULES
1550                         if (request_module("nfct-helper-%s", helpname) < 0) {
1551                                 err = -EOPNOTSUPP;
1552                                 goto err1;
1553                         }
1554
1555                         rcu_read_lock();
1556                         helper = __nf_conntrack_helper_find(helpname,
1557                                                             nf_ct_l3num(ct),
1558                                                             nf_ct_protonum(ct));
1559                         if (helper) {
1560                                 err = -EAGAIN;
1561                                 goto err2;
1562                         }
1563                         rcu_read_unlock();
1564 #endif
1565                         err = -EOPNOTSUPP;
1566                         goto err1;
1567                 } else {
1568                         struct nf_conn_help *help;
1569
1570                         help = nf_ct_helper_ext_add(ct, helper, GFP_ATOMIC);
1571                         if (help == NULL) {
1572                                 err = -ENOMEM;
1573                                 goto err2;
1574                         }
1575                         /* set private helper data if allowed. */
1576                         if (helper->from_nlattr)
1577                                 helper->from_nlattr(helpinfo, ct);
1578
1579                         /* not in hash table yet so not strictly necessary */
1580                         RCU_INIT_POINTER(help->helper, helper);
1581                 }
1582         } else {
1583                 /* try an implicit helper assignation */
1584                 err = __nf_ct_try_assign_helper(ct, NULL, GFP_ATOMIC);
1585                 if (err < 0)
1586                         goto err2;
1587         }
1588
1589         if (cda[CTA_NAT_SRC] || cda[CTA_NAT_DST]) {
1590                 err = ctnetlink_change_nat(ct, cda);
1591                 if (err < 0)
1592                         goto err2;
1593         }
1594
1595         nf_ct_acct_ext_add(ct, GFP_ATOMIC);
1596         nf_ct_tstamp_ext_add(ct, GFP_ATOMIC);
1597         nf_ct_ecache_ext_add(ct, 0, 0, GFP_ATOMIC);
1598         /* we must add conntrack extensions before confirmation. */
1599         ct->status |= IPS_CONFIRMED;
1600
1601         if (cda[CTA_STATUS]) {
1602                 err = ctnetlink_change_status(ct, cda);
1603                 if (err < 0)
1604                         goto err2;
1605         }
1606
1607 #ifdef CONFIG_NF_NAT_NEEDED
1608         if (cda[CTA_NAT_SEQ_ADJ_ORIG] || cda[CTA_NAT_SEQ_ADJ_REPLY]) {
1609                 err = ctnetlink_change_nat_seq_adj(ct, cda);
1610                 if (err < 0)
1611                         goto err2;
1612         }
1613 #endif
1614
1615         memset(&ct->proto, 0, sizeof(ct->proto));
1616         if (cda[CTA_PROTOINFO]) {
1617                 err = ctnetlink_change_protoinfo(ct, cda);
1618                 if (err < 0)
1619                         goto err2;
1620         }
1621
1622 #if defined(CONFIG_NF_CONNTRACK_MARK)
1623         if (cda[CTA_MARK])
1624                 ct->mark = ntohl(nla_get_be32(cda[CTA_MARK]));
1625 #endif
1626
1627         /* setup master conntrack: this is a confirmed expectation */
1628         if (cda[CTA_TUPLE_MASTER]) {
1629                 struct nf_conntrack_tuple master;
1630                 struct nf_conntrack_tuple_hash *master_h;
1631                 struct nf_conn *master_ct;
1632
1633                 err = ctnetlink_parse_tuple(cda, &master, CTA_TUPLE_MASTER, u3);
1634                 if (err < 0)
1635                         goto err2;
1636
1637                 master_h = nf_conntrack_find_get(net, zone, &master);
1638                 if (master_h == NULL) {
1639                         err = -ENOENT;
1640                         goto err2;
1641                 }
1642                 master_ct = nf_ct_tuplehash_to_ctrack(master_h);
1643                 __set_bit(IPS_EXPECTED_BIT, &ct->status);
1644                 ct->master = master_ct;
1645         }
1646         tstamp = nf_conn_tstamp_find(ct);
1647         if (tstamp)
1648                 tstamp->start = ktime_to_ns(ktime_get_real());
1649
1650         err = nf_conntrack_hash_check_insert(ct);
1651         if (err < 0)
1652                 goto err2;
1653
1654         rcu_read_unlock();
1655
1656         return ct;
1657
1658 err2:
1659         rcu_read_unlock();
1660 err1:
1661         nf_conntrack_free(ct);
1662         return ERR_PTR(err);
1663 }
1664
1665 static int
1666 ctnetlink_new_conntrack(struct sock *ctnl, struct sk_buff *skb,
1667                         const struct nlmsghdr *nlh,
1668                         const struct nlattr * const cda[])
1669 {
1670         struct net *net = sock_net(ctnl);
1671         struct nf_conntrack_tuple otuple, rtuple;
1672         struct nf_conntrack_tuple_hash *h = NULL;
1673         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
1674         struct nf_conn *ct;
1675         u_int8_t u3 = nfmsg->nfgen_family;
1676         u16 zone;
1677         int err;
1678
1679         err = ctnetlink_parse_zone(cda[CTA_ZONE], &zone);
1680         if (err < 0)
1681                 return err;
1682
1683         if (cda[CTA_TUPLE_ORIG]) {
1684                 err = ctnetlink_parse_tuple(cda, &otuple, CTA_TUPLE_ORIG, u3);
1685                 if (err < 0)
1686                         return err;
1687         }
1688
1689         if (cda[CTA_TUPLE_REPLY]) {
1690                 err = ctnetlink_parse_tuple(cda, &rtuple, CTA_TUPLE_REPLY, u3);
1691                 if (err < 0)
1692                         return err;
1693         }
1694
1695         if (cda[CTA_TUPLE_ORIG])
1696                 h = nf_conntrack_find_get(net, zone, &otuple);
1697         else if (cda[CTA_TUPLE_REPLY])
1698                 h = nf_conntrack_find_get(net, zone, &rtuple);
1699
1700         if (h == NULL) {
1701                 err = -ENOENT;
1702                 if (nlh->nlmsg_flags & NLM_F_CREATE) {
1703                         enum ip_conntrack_events events;
1704
1705                         ct = ctnetlink_create_conntrack(net, zone, cda, &otuple,
1706                                                         &rtuple, u3);
1707                         if (IS_ERR(ct))
1708                                 return PTR_ERR(ct);
1709
1710                         err = 0;
1711                         if (test_bit(IPS_EXPECTED_BIT, &ct->status))
1712                                 events = IPCT_RELATED;
1713                         else
1714                                 events = IPCT_NEW;
1715
1716                         nf_conntrack_eventmask_report((1 << IPCT_REPLY) |
1717                                                       (1 << IPCT_ASSURED) |
1718                                                       (1 << IPCT_HELPER) |
1719                                                       (1 << IPCT_PROTOINFO) |
1720                                                       (1 << IPCT_NATSEQADJ) |
1721                                                       (1 << IPCT_MARK) | events,
1722                                                       ct, NETLINK_CB(skb).portid,
1723                                                       nlmsg_report(nlh));
1724                         nf_ct_put(ct);
1725                 }
1726
1727                 return err;
1728         }
1729         /* implicit 'else' */
1730
1731         err = -EEXIST;
1732         ct = nf_ct_tuplehash_to_ctrack(h);
1733         if (!(nlh->nlmsg_flags & NLM_F_EXCL)) {
1734                 spin_lock_bh(&nf_conntrack_lock);
1735                 err = ctnetlink_change_conntrack(ct, cda);
1736                 spin_unlock_bh(&nf_conntrack_lock);
1737                 if (err == 0) {
1738                         nf_conntrack_eventmask_report((1 << IPCT_REPLY) |
1739                                                       (1 << IPCT_ASSURED) |
1740                                                       (1 << IPCT_HELPER) |
1741                                                       (1 << IPCT_PROTOINFO) |
1742                                                       (1 << IPCT_NATSEQADJ) |
1743                                                       (1 << IPCT_MARK),
1744                                                       ct, NETLINK_CB(skb).portid,
1745                                                       nlmsg_report(nlh));
1746                 }
1747         }
1748
1749         nf_ct_put(ct);
1750         return err;
1751 }
1752
1753 static int
1754 ctnetlink_ct_stat_cpu_fill_info(struct sk_buff *skb, u32 portid, u32 seq,
1755                                 __u16 cpu, const struct ip_conntrack_stat *st)
1756 {
1757         struct nlmsghdr *nlh;
1758         struct nfgenmsg *nfmsg;
1759         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
1760
1761         event = (NFNL_SUBSYS_CTNETLINK << 8 | IPCTNL_MSG_CT_GET_STATS_CPU);
1762         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
1763         if (nlh == NULL)
1764                 goto nlmsg_failure;
1765
1766         nfmsg = nlmsg_data(nlh);
1767         nfmsg->nfgen_family = AF_UNSPEC;
1768         nfmsg->version      = NFNETLINK_V0;
1769         nfmsg->res_id       = htons(cpu);
1770
1771         if (nla_put_be32(skb, CTA_STATS_SEARCHED, htonl(st->searched)) ||
1772             nla_put_be32(skb, CTA_STATS_FOUND, htonl(st->found)) ||
1773             nla_put_be32(skb, CTA_STATS_NEW, htonl(st->new)) ||
1774             nla_put_be32(skb, CTA_STATS_INVALID, htonl(st->invalid)) ||
1775             nla_put_be32(skb, CTA_STATS_IGNORE, htonl(st->ignore)) ||
1776             nla_put_be32(skb, CTA_STATS_DELETE, htonl(st->delete)) ||
1777             nla_put_be32(skb, CTA_STATS_DELETE_LIST, htonl(st->delete_list)) ||
1778             nla_put_be32(skb, CTA_STATS_INSERT, htonl(st->insert)) ||
1779             nla_put_be32(skb, CTA_STATS_INSERT_FAILED,
1780                                 htonl(st->insert_failed)) ||
1781             nla_put_be32(skb, CTA_STATS_DROP, htonl(st->drop)) ||
1782             nla_put_be32(skb, CTA_STATS_EARLY_DROP, htonl(st->early_drop)) ||
1783             nla_put_be32(skb, CTA_STATS_ERROR, htonl(st->error)) ||
1784             nla_put_be32(skb, CTA_STATS_SEARCH_RESTART,
1785                                 htonl(st->search_restart)))
1786                 goto nla_put_failure;
1787
1788         nlmsg_end(skb, nlh);
1789         return skb->len;
1790
1791 nla_put_failure:
1792 nlmsg_failure:
1793         nlmsg_cancel(skb, nlh);
1794         return -1;
1795 }
1796
1797 static int
1798 ctnetlink_ct_stat_cpu_dump(struct sk_buff *skb, struct netlink_callback *cb)
1799 {
1800         int cpu;
1801         struct net *net = sock_net(skb->sk);
1802
1803         if (cb->args[0] == nr_cpu_ids)
1804                 return 0;
1805
1806         for (cpu = cb->args[0]; cpu < nr_cpu_ids; cpu++) {
1807                 const struct ip_conntrack_stat *st;
1808
1809                 if (!cpu_possible(cpu))
1810                         continue;
1811
1812                 st = per_cpu_ptr(net->ct.stat, cpu);
1813                 if (ctnetlink_ct_stat_cpu_fill_info(skb,
1814                                                     NETLINK_CB(cb->skb).portid,
1815                                                     cb->nlh->nlmsg_seq,
1816                                                     cpu, st) < 0)
1817                                 break;
1818         }
1819         cb->args[0] = cpu;
1820
1821         return skb->len;
1822 }
1823
1824 static int
1825 ctnetlink_stat_ct_cpu(struct sock *ctnl, struct sk_buff *skb,
1826                       const struct nlmsghdr *nlh,
1827                       const struct nlattr * const cda[])
1828 {
1829         if (nlh->nlmsg_flags & NLM_F_DUMP) {
1830                 struct netlink_dump_control c = {
1831                         .dump = ctnetlink_ct_stat_cpu_dump,
1832                 };
1833                 return netlink_dump_start(ctnl, skb, nlh, &c);
1834         }
1835
1836         return 0;
1837 }
1838
1839 static int
1840 ctnetlink_stat_ct_fill_info(struct sk_buff *skb, u32 portid, u32 seq, u32 type,
1841                             struct net *net)
1842 {
1843         struct nlmsghdr *nlh;
1844         struct nfgenmsg *nfmsg;
1845         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
1846         unsigned int nr_conntracks = atomic_read(&net->ct.count);
1847
1848         event = (NFNL_SUBSYS_CTNETLINK << 8 | IPCTNL_MSG_CT_GET_STATS);
1849         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
1850         if (nlh == NULL)
1851                 goto nlmsg_failure;
1852
1853         nfmsg = nlmsg_data(nlh);
1854         nfmsg->nfgen_family = AF_UNSPEC;
1855         nfmsg->version      = NFNETLINK_V0;
1856         nfmsg->res_id       = 0;
1857
1858         if (nla_put_be32(skb, CTA_STATS_GLOBAL_ENTRIES, htonl(nr_conntracks)))
1859                 goto nla_put_failure;
1860
1861         nlmsg_end(skb, nlh);
1862         return skb->len;
1863
1864 nla_put_failure:
1865 nlmsg_failure:
1866         nlmsg_cancel(skb, nlh);
1867         return -1;
1868 }
1869
1870 static int
1871 ctnetlink_stat_ct(struct sock *ctnl, struct sk_buff *skb,
1872                   const struct nlmsghdr *nlh,
1873                   const struct nlattr * const cda[])
1874 {
1875         struct sk_buff *skb2;
1876         int err;
1877
1878         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
1879         if (skb2 == NULL)
1880                 return -ENOMEM;
1881
1882         err = ctnetlink_stat_ct_fill_info(skb2, NETLINK_CB(skb).portid,
1883                                           nlh->nlmsg_seq,
1884                                           NFNL_MSG_TYPE(nlh->nlmsg_type),
1885                                           sock_net(skb->sk));
1886         if (err <= 0)
1887                 goto free;
1888
1889         err = netlink_unicast(ctnl, skb2, NETLINK_CB(skb).portid, MSG_DONTWAIT);
1890         if (err < 0)
1891                 goto out;
1892
1893         return 0;
1894
1895 free:
1896         kfree_skb(skb2);
1897 out:
1898         /* this avoids a loop in nfnetlink. */
1899         return err == -EAGAIN ? -ENOBUFS : err;
1900 }
1901
1902 #ifdef CONFIG_NETFILTER_NETLINK_QUEUE_CT
1903 static size_t
1904 ctnetlink_nfqueue_build_size(const struct nf_conn *ct)
1905 {
1906         return 3 * nla_total_size(0) /* CTA_TUPLE_ORIG|REPL|MASTER */
1907                + 3 * nla_total_size(0) /* CTA_TUPLE_IP */
1908                + 3 * nla_total_size(0) /* CTA_TUPLE_PROTO */
1909                + 3 * nla_total_size(sizeof(u_int8_t)) /* CTA_PROTO_NUM */
1910                + nla_total_size(sizeof(u_int32_t)) /* CTA_ID */
1911                + nla_total_size(sizeof(u_int32_t)) /* CTA_STATUS */
1912                + nla_total_size(sizeof(u_int32_t)) /* CTA_TIMEOUT */
1913                + nla_total_size(0) /* CTA_PROTOINFO */
1914                + nla_total_size(0) /* CTA_HELP */
1915                + nla_total_size(NF_CT_HELPER_NAME_LEN) /* CTA_HELP_NAME */
1916                + ctnetlink_secctx_size(ct)
1917 #ifdef CONFIG_NF_NAT_NEEDED
1918                + 2 * nla_total_size(0) /* CTA_NAT_SEQ_ADJ_ORIG|REPL */
1919                + 6 * nla_total_size(sizeof(u_int32_t)) /* CTA_NAT_SEQ_OFFSET */
1920 #endif
1921 #ifdef CONFIG_NF_CONNTRACK_MARK
1922                + nla_total_size(sizeof(u_int32_t)) /* CTA_MARK */
1923 #endif
1924                + ctnetlink_proto_size(ct)
1925                ;
1926 }
1927
1928 static int
1929 ctnetlink_nfqueue_build(struct sk_buff *skb, struct nf_conn *ct)
1930 {
1931         struct nlattr *nest_parms;
1932
1933         rcu_read_lock();
1934         nest_parms = nla_nest_start(skb, CTA_TUPLE_ORIG | NLA_F_NESTED);
1935         if (!nest_parms)
1936                 goto nla_put_failure;
1937         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_ORIGINAL)) < 0)
1938                 goto nla_put_failure;
1939         nla_nest_end(skb, nest_parms);
1940
1941         nest_parms = nla_nest_start(skb, CTA_TUPLE_REPLY | NLA_F_NESTED);
1942         if (!nest_parms)
1943                 goto nla_put_failure;
1944         if (ctnetlink_dump_tuples(skb, nf_ct_tuple(ct, IP_CT_DIR_REPLY)) < 0)
1945                 goto nla_put_failure;
1946         nla_nest_end(skb, nest_parms);
1947
1948         if (nf_ct_zone(ct)) {
1949                 if (nla_put_be16(skb, CTA_ZONE, htons(nf_ct_zone(ct))))
1950                         goto nla_put_failure;
1951         }
1952
1953         if (ctnetlink_dump_id(skb, ct) < 0)
1954                 goto nla_put_failure;
1955
1956         if (ctnetlink_dump_status(skb, ct) < 0)
1957                 goto nla_put_failure;
1958
1959         if (ctnetlink_dump_timeout(skb, ct) < 0)
1960                 goto nla_put_failure;
1961
1962         if (ctnetlink_dump_protoinfo(skb, ct) < 0)
1963                 goto nla_put_failure;
1964
1965         if (ctnetlink_dump_helpinfo(skb, ct) < 0)
1966                 goto nla_put_failure;
1967
1968 #ifdef CONFIG_NF_CONNTRACK_SECMARK
1969         if (ct->secmark && ctnetlink_dump_secctx(skb, ct) < 0)
1970                 goto nla_put_failure;
1971 #endif
1972         if (ct->master && ctnetlink_dump_master(skb, ct) < 0)
1973                 goto nla_put_failure;
1974
1975         if ((ct->status & IPS_SEQ_ADJUST) &&
1976             ctnetlink_dump_nat_seq_adj(skb, ct) < 0)
1977                 goto nla_put_failure;
1978
1979 #ifdef CONFIG_NF_CONNTRACK_MARK
1980         if (ct->mark && ctnetlink_dump_mark(skb, ct) < 0)
1981                 goto nla_put_failure;
1982 #endif
1983         rcu_read_unlock();
1984         return 0;
1985
1986 nla_put_failure:
1987         rcu_read_unlock();
1988         return -ENOSPC;
1989 }
1990
1991 static int
1992 ctnetlink_nfqueue_parse_ct(const struct nlattr *cda[], struct nf_conn *ct)
1993 {
1994         int err;
1995
1996         if (cda[CTA_TIMEOUT]) {
1997                 err = ctnetlink_change_timeout(ct, cda);
1998                 if (err < 0)
1999                         return err;
2000         }
2001         if (cda[CTA_STATUS]) {
2002                 err = ctnetlink_change_status(ct, cda);
2003                 if (err < 0)
2004                         return err;
2005         }
2006         if (cda[CTA_HELP]) {
2007                 err = ctnetlink_change_helper(ct, cda);
2008                 if (err < 0)
2009                         return err;
2010         }
2011 #if defined(CONFIG_NF_CONNTRACK_MARK)
2012         if (cda[CTA_MARK])
2013                 ct->mark = ntohl(nla_get_be32(cda[CTA_MARK]));
2014 #endif
2015         return 0;
2016 }
2017
2018 static int
2019 ctnetlink_nfqueue_parse(const struct nlattr *attr, struct nf_conn *ct)
2020 {
2021         struct nlattr *cda[CTA_MAX+1];
2022         int ret;
2023
2024         nla_parse_nested(cda, CTA_MAX, attr, ct_nla_policy);
2025
2026         spin_lock_bh(&nf_conntrack_lock);
2027         ret = ctnetlink_nfqueue_parse_ct((const struct nlattr **)cda, ct);
2028         spin_unlock_bh(&nf_conntrack_lock);
2029
2030         return ret;
2031 }
2032
2033 static struct nfq_ct_hook ctnetlink_nfqueue_hook = {
2034         .build_size     = ctnetlink_nfqueue_build_size,
2035         .build          = ctnetlink_nfqueue_build,
2036         .parse          = ctnetlink_nfqueue_parse,
2037 };
2038 #endif /* CONFIG_NETFILTER_NETLINK_QUEUE_CT */
2039
2040 /***********************************************************************
2041  * EXPECT
2042  ***********************************************************************/
2043
2044 static inline int
2045 ctnetlink_exp_dump_tuple(struct sk_buff *skb,
2046                          const struct nf_conntrack_tuple *tuple,
2047                          enum ctattr_expect type)
2048 {
2049         struct nlattr *nest_parms;
2050
2051         nest_parms = nla_nest_start(skb, type | NLA_F_NESTED);
2052         if (!nest_parms)
2053                 goto nla_put_failure;
2054         if (ctnetlink_dump_tuples(skb, tuple) < 0)
2055                 goto nla_put_failure;
2056         nla_nest_end(skb, nest_parms);
2057
2058         return 0;
2059
2060 nla_put_failure:
2061         return -1;
2062 }
2063
2064 static inline int
2065 ctnetlink_exp_dump_mask(struct sk_buff *skb,
2066                         const struct nf_conntrack_tuple *tuple,
2067                         const struct nf_conntrack_tuple_mask *mask)
2068 {
2069         int ret;
2070         struct nf_conntrack_l3proto *l3proto;
2071         struct nf_conntrack_l4proto *l4proto;
2072         struct nf_conntrack_tuple m;
2073         struct nlattr *nest_parms;
2074
2075         memset(&m, 0xFF, sizeof(m));
2076         memcpy(&m.src.u3, &mask->src.u3, sizeof(m.src.u3));
2077         m.src.u.all = mask->src.u.all;
2078         m.dst.protonum = tuple->dst.protonum;
2079
2080         nest_parms = nla_nest_start(skb, CTA_EXPECT_MASK | NLA_F_NESTED);
2081         if (!nest_parms)
2082                 goto nla_put_failure;
2083
2084         rcu_read_lock();
2085         l3proto = __nf_ct_l3proto_find(tuple->src.l3num);
2086         ret = ctnetlink_dump_tuples_ip(skb, &m, l3proto);
2087         if (ret >= 0) {
2088                 l4proto = __nf_ct_l4proto_find(tuple->src.l3num,
2089                                                tuple->dst.protonum);
2090         ret = ctnetlink_dump_tuples_proto(skb, &m, l4proto);
2091         }
2092         rcu_read_unlock();
2093
2094         if (unlikely(ret < 0))
2095                 goto nla_put_failure;
2096
2097         nla_nest_end(skb, nest_parms);
2098
2099         return 0;
2100
2101 nla_put_failure:
2102         return -1;
2103 }
2104
2105 static const union nf_inet_addr any_addr;
2106
2107 static int
2108 ctnetlink_exp_dump_expect(struct sk_buff *skb,
2109                           const struct nf_conntrack_expect *exp)
2110 {
2111         struct nf_conn *master = exp->master;
2112         long timeout = ((long)exp->timeout.expires - (long)jiffies) / HZ;
2113         struct nf_conn_help *help;
2114 #ifdef CONFIG_NF_NAT_NEEDED
2115         struct nlattr *nest_parms;
2116         struct nf_conntrack_tuple nat_tuple = {};
2117 #endif
2118         struct nf_ct_helper_expectfn *expfn;
2119
2120         if (timeout < 0)
2121                 timeout = 0;
2122
2123         if (ctnetlink_exp_dump_tuple(skb, &exp->tuple, CTA_EXPECT_TUPLE) < 0)
2124                 goto nla_put_failure;
2125         if (ctnetlink_exp_dump_mask(skb, &exp->tuple, &exp->mask) < 0)
2126                 goto nla_put_failure;
2127         if (ctnetlink_exp_dump_tuple(skb,
2128                                  &master->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
2129                                  CTA_EXPECT_MASTER) < 0)
2130                 goto nla_put_failure;
2131
2132 #ifdef CONFIG_NF_NAT_NEEDED
2133         if (!nf_inet_addr_cmp(&exp->saved_addr, &any_addr) ||
2134             exp->saved_proto.all) {
2135                 nest_parms = nla_nest_start(skb, CTA_EXPECT_NAT | NLA_F_NESTED);
2136                 if (!nest_parms)
2137                         goto nla_put_failure;
2138
2139                 if (nla_put_be32(skb, CTA_EXPECT_NAT_DIR, htonl(exp->dir)))
2140                         goto nla_put_failure;
2141
2142                 nat_tuple.src.l3num = nf_ct_l3num(master);
2143                 nat_tuple.src.u3 = exp->saved_addr;
2144                 nat_tuple.dst.protonum = nf_ct_protonum(master);
2145                 nat_tuple.src.u = exp->saved_proto;
2146
2147                 if (ctnetlink_exp_dump_tuple(skb, &nat_tuple,
2148                                                 CTA_EXPECT_NAT_TUPLE) < 0)
2149                         goto nla_put_failure;
2150                 nla_nest_end(skb, nest_parms);
2151         }
2152 #endif
2153         if (nla_put_be32(skb, CTA_EXPECT_TIMEOUT, htonl(timeout)) ||
2154             nla_put_be32(skb, CTA_EXPECT_ID, htonl((unsigned long)exp)) ||
2155             nla_put_be32(skb, CTA_EXPECT_FLAGS, htonl(exp->flags)) ||
2156             nla_put_be32(skb, CTA_EXPECT_CLASS, htonl(exp->class)))
2157                 goto nla_put_failure;
2158         help = nfct_help(master);
2159         if (help) {
2160                 struct nf_conntrack_helper *helper;
2161
2162                 helper = rcu_dereference(help->helper);
2163                 if (helper &&
2164                     nla_put_string(skb, CTA_EXPECT_HELP_NAME, helper->name))
2165                         goto nla_put_failure;
2166         }
2167         expfn = nf_ct_helper_expectfn_find_by_symbol(exp->expectfn);
2168         if (expfn != NULL &&
2169             nla_put_string(skb, CTA_EXPECT_FN, expfn->name))
2170                 goto nla_put_failure;
2171
2172         return 0;
2173
2174 nla_put_failure:
2175         return -1;
2176 }
2177
2178 static int
2179 ctnetlink_exp_fill_info(struct sk_buff *skb, u32 portid, u32 seq,
2180                         int event, const struct nf_conntrack_expect *exp)
2181 {
2182         struct nlmsghdr *nlh;
2183         struct nfgenmsg *nfmsg;
2184         unsigned int flags = portid ? NLM_F_MULTI : 0;
2185
2186         event |= NFNL_SUBSYS_CTNETLINK_EXP << 8;
2187         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
2188         if (nlh == NULL)
2189                 goto nlmsg_failure;
2190
2191         nfmsg = nlmsg_data(nlh);
2192         nfmsg->nfgen_family = exp->tuple.src.l3num;
2193         nfmsg->version      = NFNETLINK_V0;
2194         nfmsg->res_id       = 0;
2195
2196         if (ctnetlink_exp_dump_expect(skb, exp) < 0)
2197                 goto nla_put_failure;
2198
2199         nlmsg_end(skb, nlh);
2200         return skb->len;
2201
2202 nlmsg_failure:
2203 nla_put_failure:
2204         nlmsg_cancel(skb, nlh);
2205         return -1;
2206 }
2207
2208 #ifdef CONFIG_NF_CONNTRACK_EVENTS
2209 static int
2210 ctnetlink_expect_event(unsigned int events, struct nf_exp_event *item)
2211 {
2212         struct nf_conntrack_expect *exp = item->exp;
2213         struct net *net = nf_ct_exp_net(exp);
2214         struct nlmsghdr *nlh;
2215         struct nfgenmsg *nfmsg;
2216         struct sk_buff *skb;
2217         unsigned int type, group;
2218         int flags = 0;
2219
2220         if (events & (1 << IPEXP_DESTROY)) {
2221                 type = IPCTNL_MSG_EXP_DELETE;
2222                 group = NFNLGRP_CONNTRACK_EXP_DESTROY;
2223         } else if (events & (1 << IPEXP_NEW)) {
2224                 type = IPCTNL_MSG_EXP_NEW;
2225                 flags = NLM_F_CREATE|NLM_F_EXCL;
2226                 group = NFNLGRP_CONNTRACK_EXP_NEW;
2227         } else
2228                 return 0;
2229
2230         if (!item->report && !nfnetlink_has_listeners(net, group))
2231                 return 0;
2232
2233         skb = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_ATOMIC);
2234         if (skb == NULL)
2235                 goto errout;
2236
2237         type |= NFNL_SUBSYS_CTNETLINK_EXP << 8;
2238         nlh = nlmsg_put(skb, item->portid, 0, type, sizeof(*nfmsg), flags);
2239         if (nlh == NULL)
2240                 goto nlmsg_failure;
2241
2242         nfmsg = nlmsg_data(nlh);
2243         nfmsg->nfgen_family = exp->tuple.src.l3num;
2244         nfmsg->version      = NFNETLINK_V0;
2245         nfmsg->res_id       = 0;
2246
2247         rcu_read_lock();
2248         if (ctnetlink_exp_dump_expect(skb, exp) < 0)
2249                 goto nla_put_failure;
2250         rcu_read_unlock();
2251
2252         nlmsg_end(skb, nlh);
2253         nfnetlink_send(skb, net, item->portid, group, item->report, GFP_ATOMIC);
2254         return 0;
2255
2256 nla_put_failure:
2257         rcu_read_unlock();
2258         nlmsg_cancel(skb, nlh);
2259 nlmsg_failure:
2260         kfree_skb(skb);
2261 errout:
2262         nfnetlink_set_err(net, 0, 0, -ENOBUFS);
2263         return 0;
2264 }
2265 #endif
2266 static int ctnetlink_exp_done(struct netlink_callback *cb)
2267 {
2268         if (cb->args[1])
2269                 nf_ct_expect_put((struct nf_conntrack_expect *)cb->args[1]);
2270         return 0;
2271 }
2272
2273 static int
2274 ctnetlink_exp_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
2275 {
2276         struct net *net = sock_net(skb->sk);
2277         struct nf_conntrack_expect *exp, *last;
2278         struct nfgenmsg *nfmsg = nlmsg_data(cb->nlh);
2279         struct hlist_node *n;
2280         u_int8_t l3proto = nfmsg->nfgen_family;
2281
2282         rcu_read_lock();
2283         last = (struct nf_conntrack_expect *)cb->args[1];
2284         for (; cb->args[0] < nf_ct_expect_hsize; cb->args[0]++) {
2285 restart:
2286                 hlist_for_each_entry(exp, n, &net->ct.expect_hash[cb->args[0]],
2287                                      hnode) {
2288                         if (l3proto && exp->tuple.src.l3num != l3proto)
2289                                 continue;
2290                         if (cb->args[1]) {
2291                                 if (exp != last)
2292                                         continue;
2293                                 cb->args[1] = 0;
2294                         }
2295                         if (ctnetlink_exp_fill_info(skb,
2296                                                     NETLINK_CB(cb->skb).portid,
2297                                                     cb->nlh->nlmsg_seq,
2298                                                     IPCTNL_MSG_EXP_NEW,
2299                                                     exp) < 0) {
2300                                 if (!atomic_inc_not_zero(&exp->use))
2301                                         continue;
2302                                 cb->args[1] = (unsigned long)exp;
2303                                 goto out;
2304                         }
2305                 }
2306                 if (cb->args[1]) {
2307                         cb->args[1] = 0;
2308                         goto restart;
2309                 }
2310         }
2311 out:
2312         rcu_read_unlock();
2313         if (last)
2314                 nf_ct_expect_put(last);
2315
2316         return skb->len;
2317 }
2318
2319 static const struct nla_policy exp_nla_policy[CTA_EXPECT_MAX+1] = {
2320         [CTA_EXPECT_MASTER]     = { .type = NLA_NESTED },
2321         [CTA_EXPECT_TUPLE]      = { .type = NLA_NESTED },
2322         [CTA_EXPECT_MASK]       = { .type = NLA_NESTED },
2323         [CTA_EXPECT_TIMEOUT]    = { .type = NLA_U32 },
2324         [CTA_EXPECT_ID]         = { .type = NLA_U32 },
2325         [CTA_EXPECT_HELP_NAME]  = { .type = NLA_NUL_STRING },
2326         [CTA_EXPECT_ZONE]       = { .type = NLA_U16 },
2327         [CTA_EXPECT_FLAGS]      = { .type = NLA_U32 },
2328         [CTA_EXPECT_CLASS]      = { .type = NLA_U32 },
2329         [CTA_EXPECT_NAT]        = { .type = NLA_NESTED },
2330         [CTA_EXPECT_FN]         = { .type = NLA_NUL_STRING },
2331 };
2332
2333 static int
2334 ctnetlink_get_expect(struct sock *ctnl, struct sk_buff *skb,
2335                      const struct nlmsghdr *nlh,
2336                      const struct nlattr * const cda[])
2337 {
2338         struct net *net = sock_net(ctnl);
2339         struct nf_conntrack_tuple tuple;
2340         struct nf_conntrack_expect *exp;
2341         struct sk_buff *skb2;
2342         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
2343         u_int8_t u3 = nfmsg->nfgen_family;
2344         u16 zone;
2345         int err;
2346
2347         if (nlh->nlmsg_flags & NLM_F_DUMP) {
2348                 struct netlink_dump_control c = {
2349                         .dump = ctnetlink_exp_dump_table,
2350                         .done = ctnetlink_exp_done,
2351                 };
2352                 return netlink_dump_start(ctnl, skb, nlh, &c);
2353         }
2354
2355         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
2356         if (err < 0)
2357                 return err;
2358
2359         if (cda[CTA_EXPECT_TUPLE])
2360                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE, u3);
2361         else if (cda[CTA_EXPECT_MASTER])
2362                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_MASTER, u3);
2363         else
2364                 return -EINVAL;
2365
2366         if (err < 0)
2367                 return err;
2368
2369         exp = nf_ct_expect_find_get(net, zone, &tuple);
2370         if (!exp)
2371                 return -ENOENT;
2372
2373         if (cda[CTA_EXPECT_ID]) {
2374                 __be32 id = nla_get_be32(cda[CTA_EXPECT_ID]);
2375                 if (ntohl(id) != (u32)(unsigned long)exp) {
2376                         nf_ct_expect_put(exp);
2377                         return -ENOENT;
2378                 }
2379         }
2380
2381         err = -ENOMEM;
2382         skb2 = nlmsg_new(NLMSG_DEFAULT_SIZE, GFP_KERNEL);
2383         if (skb2 == NULL) {
2384                 nf_ct_expect_put(exp);
2385                 goto out;
2386         }
2387
2388         rcu_read_lock();
2389         err = ctnetlink_exp_fill_info(skb2, NETLINK_CB(skb).portid,
2390                                       nlh->nlmsg_seq, IPCTNL_MSG_EXP_NEW, exp);
2391         rcu_read_unlock();
2392         nf_ct_expect_put(exp);
2393         if (err <= 0)
2394                 goto free;
2395
2396         err = netlink_unicast(ctnl, skb2, NETLINK_CB(skb).portid, MSG_DONTWAIT);
2397         if (err < 0)
2398                 goto out;
2399
2400         return 0;
2401
2402 free:
2403         kfree_skb(skb2);
2404 out:
2405         /* this avoids a loop in nfnetlink. */
2406         return err == -EAGAIN ? -ENOBUFS : err;
2407 }
2408
2409 static int
2410 ctnetlink_del_expect(struct sock *ctnl, struct sk_buff *skb,
2411                      const struct nlmsghdr *nlh,
2412                      const struct nlattr * const cda[])
2413 {
2414         struct net *net = sock_net(ctnl);
2415         struct nf_conntrack_expect *exp;
2416         struct nf_conntrack_tuple tuple;
2417         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
2418         struct hlist_node *n, *next;
2419         u_int8_t u3 = nfmsg->nfgen_family;
2420         unsigned int i;
2421         u16 zone;
2422         int err;
2423
2424         if (cda[CTA_EXPECT_TUPLE]) {
2425                 /* delete a single expect by tuple */
2426                 err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
2427                 if (err < 0)
2428                         return err;
2429
2430                 err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE, u3);
2431                 if (err < 0)
2432                         return err;
2433
2434                 /* bump usage count to 2 */
2435                 exp = nf_ct_expect_find_get(net, zone, &tuple);
2436                 if (!exp)
2437                         return -ENOENT;
2438
2439                 if (cda[CTA_EXPECT_ID]) {
2440                         __be32 id = nla_get_be32(cda[CTA_EXPECT_ID]);
2441                         if (ntohl(id) != (u32)(unsigned long)exp) {
2442                                 nf_ct_expect_put(exp);
2443                                 return -ENOENT;
2444                         }
2445                 }
2446
2447                 /* after list removal, usage count == 1 */
2448                 spin_lock_bh(&nf_conntrack_lock);
2449                 if (del_timer(&exp->timeout)) {
2450                         nf_ct_unlink_expect_report(exp, NETLINK_CB(skb).portid,
2451                                                    nlmsg_report(nlh));
2452                         nf_ct_expect_put(exp);
2453                 }
2454                 spin_unlock_bh(&nf_conntrack_lock);
2455                 /* have to put what we 'get' above.
2456                  * after this line usage count == 0 */
2457                 nf_ct_expect_put(exp);
2458         } else if (cda[CTA_EXPECT_HELP_NAME]) {
2459                 char *name = nla_data(cda[CTA_EXPECT_HELP_NAME]);
2460                 struct nf_conn_help *m_help;
2461
2462                 /* delete all expectations for this helper */
2463                 spin_lock_bh(&nf_conntrack_lock);
2464                 for (i = 0; i < nf_ct_expect_hsize; i++) {
2465                         hlist_for_each_entry_safe(exp, n, next,
2466                                                   &net->ct.expect_hash[i],
2467                                                   hnode) {
2468                                 m_help = nfct_help(exp->master);
2469                                 if (!strcmp(m_help->helper->name, name) &&
2470                                     del_timer(&exp->timeout)) {
2471                                         nf_ct_unlink_expect_report(exp,
2472                                                         NETLINK_CB(skb).portid,
2473                                                         nlmsg_report(nlh));
2474                                         nf_ct_expect_put(exp);
2475                                 }
2476                         }
2477                 }
2478                 spin_unlock_bh(&nf_conntrack_lock);
2479         } else {
2480                 /* This basically means we have to flush everything*/
2481                 spin_lock_bh(&nf_conntrack_lock);
2482                 for (i = 0; i < nf_ct_expect_hsize; i++) {
2483                         hlist_for_each_entry_safe(exp, n, next,
2484                                                   &net->ct.expect_hash[i],
2485                                                   hnode) {
2486                                 if (del_timer(&exp->timeout)) {
2487                                         nf_ct_unlink_expect_report(exp,
2488                                                         NETLINK_CB(skb).portid,
2489                                                         nlmsg_report(nlh));
2490                                         nf_ct_expect_put(exp);
2491                                 }
2492                         }
2493                 }
2494                 spin_unlock_bh(&nf_conntrack_lock);
2495         }
2496
2497         return 0;
2498 }
2499 static int
2500 ctnetlink_change_expect(struct nf_conntrack_expect *x,
2501                         const struct nlattr * const cda[])
2502 {
2503         if (cda[CTA_EXPECT_TIMEOUT]) {
2504                 if (!del_timer(&x->timeout))
2505                         return -ETIME;
2506
2507                 x->timeout.expires = jiffies +
2508                         ntohl(nla_get_be32(cda[CTA_EXPECT_TIMEOUT])) * HZ;
2509                 add_timer(&x->timeout);
2510         }
2511         return 0;
2512 }
2513
2514 static const struct nla_policy exp_nat_nla_policy[CTA_EXPECT_NAT_MAX+1] = {
2515         [CTA_EXPECT_NAT_DIR]    = { .type = NLA_U32 },
2516         [CTA_EXPECT_NAT_TUPLE]  = { .type = NLA_NESTED },
2517 };
2518
2519 static int
2520 ctnetlink_parse_expect_nat(const struct nlattr *attr,
2521                            struct nf_conntrack_expect *exp,
2522                            u_int8_t u3)
2523 {
2524 #ifdef CONFIG_NF_NAT_NEEDED
2525         struct nlattr *tb[CTA_EXPECT_NAT_MAX+1];
2526         struct nf_conntrack_tuple nat_tuple = {};
2527         int err;
2528
2529         nla_parse_nested(tb, CTA_EXPECT_NAT_MAX, attr, exp_nat_nla_policy);
2530
2531         if (!tb[CTA_EXPECT_NAT_DIR] || !tb[CTA_EXPECT_NAT_TUPLE])
2532                 return -EINVAL;
2533
2534         err = ctnetlink_parse_tuple((const struct nlattr * const *)tb,
2535                                         &nat_tuple, CTA_EXPECT_NAT_TUPLE, u3);
2536         if (err < 0)
2537                 return err;
2538
2539         exp->saved_addr = nat_tuple.src.u3;
2540         exp->saved_proto = nat_tuple.src.u;
2541         exp->dir = ntohl(nla_get_be32(tb[CTA_EXPECT_NAT_DIR]));
2542
2543         return 0;
2544 #else
2545         return -EOPNOTSUPP;
2546 #endif
2547 }
2548
2549 static int
2550 ctnetlink_create_expect(struct net *net, u16 zone,
2551                         const struct nlattr * const cda[],
2552                         u_int8_t u3,
2553                         u32 portid, int report)
2554 {
2555         struct nf_conntrack_tuple tuple, mask, master_tuple;
2556         struct nf_conntrack_tuple_hash *h = NULL;
2557         struct nf_conntrack_expect *exp;
2558         struct nf_conn *ct;
2559         struct nf_conn_help *help;
2560         struct nf_conntrack_helper *helper = NULL;
2561         u_int32_t class = 0;
2562         int err = 0;
2563
2564         /* caller guarantees that those three CTA_EXPECT_* exist */
2565         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE, u3);
2566         if (err < 0)
2567                 return err;
2568         err = ctnetlink_parse_tuple(cda, &mask, CTA_EXPECT_MASK, u3);
2569         if (err < 0)
2570                 return err;
2571         err = ctnetlink_parse_tuple(cda, &master_tuple, CTA_EXPECT_MASTER, u3);
2572         if (err < 0)
2573                 return err;
2574
2575         /* Look for master conntrack of this expectation */
2576         h = nf_conntrack_find_get(net, zone, &master_tuple);
2577         if (!h)
2578                 return -ENOENT;
2579         ct = nf_ct_tuplehash_to_ctrack(h);
2580
2581         /* Look for helper of this expectation */
2582         if (cda[CTA_EXPECT_HELP_NAME]) {
2583                 const char *helpname = nla_data(cda[CTA_EXPECT_HELP_NAME]);
2584
2585                 helper = __nf_conntrack_helper_find(helpname, nf_ct_l3num(ct),
2586                                                     nf_ct_protonum(ct));
2587                 if (helper == NULL) {
2588 #ifdef CONFIG_MODULES
2589                         if (request_module("nfct-helper-%s", helpname) < 0) {
2590                                 err = -EOPNOTSUPP;
2591                                 goto out;
2592                         }
2593
2594                         helper = __nf_conntrack_helper_find(helpname,
2595                                                             nf_ct_l3num(ct),
2596                                                             nf_ct_protonum(ct));
2597                         if (helper) {
2598                                 err = -EAGAIN;
2599                                 goto out;
2600                         }
2601 #endif
2602                         err = -EOPNOTSUPP;
2603                         goto out;
2604                 }
2605         }
2606
2607         if (cda[CTA_EXPECT_CLASS] && helper) {
2608                 class = ntohl(nla_get_be32(cda[CTA_EXPECT_CLASS]));
2609                 if (class > helper->expect_class_max) {
2610                         err = -EINVAL;
2611                         goto out;
2612                 }
2613         }
2614         exp = nf_ct_expect_alloc(ct);
2615         if (!exp) {
2616                 err = -ENOMEM;
2617                 goto out;
2618         }
2619         help = nfct_help(ct);
2620         if (!help) {
2621                 if (!cda[CTA_EXPECT_TIMEOUT]) {
2622                         err = -EINVAL;
2623                         goto out;
2624                 }
2625                 exp->timeout.expires =
2626                   jiffies + ntohl(nla_get_be32(cda[CTA_EXPECT_TIMEOUT])) * HZ;
2627
2628                 exp->flags = NF_CT_EXPECT_USERSPACE;
2629                 if (cda[CTA_EXPECT_FLAGS]) {
2630                         exp->flags |=
2631                                 ntohl(nla_get_be32(cda[CTA_EXPECT_FLAGS]));
2632                 }
2633         } else {
2634                 if (cda[CTA_EXPECT_FLAGS]) {
2635                         exp->flags = ntohl(nla_get_be32(cda[CTA_EXPECT_FLAGS]));
2636                         exp->flags &= ~NF_CT_EXPECT_USERSPACE;
2637                 } else
2638                         exp->flags = 0;
2639         }
2640         if (cda[CTA_EXPECT_FN]) {
2641                 const char *name = nla_data(cda[CTA_EXPECT_FN]);
2642                 struct nf_ct_helper_expectfn *expfn;
2643
2644                 expfn = nf_ct_helper_expectfn_find_by_name(name);
2645                 if (expfn == NULL) {
2646                         err = -EINVAL;
2647                         goto err_out;
2648                 }
2649                 exp->expectfn = expfn->expectfn;
2650         } else
2651                 exp->expectfn = NULL;
2652
2653         exp->class = class;
2654         exp->master = ct;
2655         exp->helper = helper;
2656         memcpy(&exp->tuple, &tuple, sizeof(struct nf_conntrack_tuple));
2657         memcpy(&exp->mask.src.u3, &mask.src.u3, sizeof(exp->mask.src.u3));
2658         exp->mask.src.u.all = mask.src.u.all;
2659
2660         if (cda[CTA_EXPECT_NAT]) {
2661                 err = ctnetlink_parse_expect_nat(cda[CTA_EXPECT_NAT],
2662                                                  exp, u3);
2663                 if (err < 0)
2664                         goto err_out;
2665         }
2666         err = nf_ct_expect_related_report(exp, portid, report);
2667 err_out:
2668         nf_ct_expect_put(exp);
2669 out:
2670         nf_ct_put(nf_ct_tuplehash_to_ctrack(h));
2671         return err;
2672 }
2673
2674 static int
2675 ctnetlink_new_expect(struct sock *ctnl, struct sk_buff *skb,
2676                      const struct nlmsghdr *nlh,
2677                      const struct nlattr * const cda[])
2678 {
2679         struct net *net = sock_net(ctnl);
2680         struct nf_conntrack_tuple tuple;
2681         struct nf_conntrack_expect *exp;
2682         struct nfgenmsg *nfmsg = nlmsg_data(nlh);
2683         u_int8_t u3 = nfmsg->nfgen_family;
2684         u16 zone;
2685         int err;
2686
2687         if (!cda[CTA_EXPECT_TUPLE]
2688             || !cda[CTA_EXPECT_MASK]
2689             || !cda[CTA_EXPECT_MASTER])
2690                 return -EINVAL;
2691
2692         err = ctnetlink_parse_zone(cda[CTA_EXPECT_ZONE], &zone);
2693         if (err < 0)
2694                 return err;
2695
2696         err = ctnetlink_parse_tuple(cda, &tuple, CTA_EXPECT_TUPLE, u3);
2697         if (err < 0)
2698                 return err;
2699
2700         spin_lock_bh(&nf_conntrack_lock);
2701         exp = __nf_ct_expect_find(net, zone, &tuple);
2702
2703         if (!exp) {
2704                 spin_unlock_bh(&nf_conntrack_lock);
2705                 err = -ENOENT;
2706                 if (nlh->nlmsg_flags & NLM_F_CREATE) {
2707                         err = ctnetlink_create_expect(net, zone, cda,
2708                                                       u3,
2709                                                       NETLINK_CB(skb).portid,
2710                                                       nlmsg_report(nlh));
2711                 }
2712                 return err;
2713         }
2714
2715         err = -EEXIST;
2716         if (!(nlh->nlmsg_flags & NLM_F_EXCL))
2717                 err = ctnetlink_change_expect(exp, cda);
2718         spin_unlock_bh(&nf_conntrack_lock);
2719
2720         return err;
2721 }
2722
2723 static int
2724 ctnetlink_exp_stat_fill_info(struct sk_buff *skb, u32 portid, u32 seq, int cpu,
2725                              const struct ip_conntrack_stat *st)
2726 {
2727         struct nlmsghdr *nlh;
2728         struct nfgenmsg *nfmsg;
2729         unsigned int flags = portid ? NLM_F_MULTI : 0, event;
2730
2731         event = (NFNL_SUBSYS_CTNETLINK << 8 | IPCTNL_MSG_EXP_GET_STATS_CPU);
2732         nlh = nlmsg_put(skb, portid, seq, event, sizeof(*nfmsg), flags);
2733         if (nlh == NULL)
2734                 goto nlmsg_failure;
2735
2736         nfmsg = nlmsg_data(nlh);
2737         nfmsg->nfgen_family = AF_UNSPEC;
2738         nfmsg->version      = NFNETLINK_V0;
2739         nfmsg->res_id       = htons(cpu);
2740
2741         if (nla_put_be32(skb, CTA_STATS_EXP_NEW, htonl(st->expect_new)) ||
2742             nla_put_be32(skb, CTA_STATS_EXP_CREATE, htonl(st->expect_create)) ||
2743             nla_put_be32(skb, CTA_STATS_EXP_DELETE, htonl(st->expect_delete)))
2744                 goto nla_put_failure;
2745
2746         nlmsg_end(skb, nlh);
2747         return skb->len;
2748
2749 nla_put_failure:
2750 nlmsg_failure:
2751         nlmsg_cancel(skb, nlh);
2752         return -1;
2753 }
2754
2755 static int
2756 ctnetlink_exp_stat_cpu_dump(struct sk_buff *skb, struct netlink_callback *cb)
2757 {
2758         int cpu;
2759         struct net *net = sock_net(skb->sk);
2760
2761         if (cb->args[0] == nr_cpu_ids)
2762                 return 0;
2763
2764         for (cpu = cb->args[0]; cpu < nr_cpu_ids; cpu++) {
2765                 const struct ip_conntrack_stat *st;
2766
2767                 if (!cpu_possible(cpu))
2768                         continue;
2769
2770                 st = per_cpu_ptr(net->ct.stat, cpu);
2771                 if (ctnetlink_exp_stat_fill_info(skb, NETLINK_CB(cb->skb).portid,
2772                                                  cb->nlh->nlmsg_seq,
2773                                                  cpu, st) < 0)
2774                         break;
2775         }
2776         cb->args[0] = cpu;
2777
2778         return skb->len;
2779 }
2780
2781 static int
2782 ctnetlink_stat_exp_cpu(struct sock *ctnl, struct sk_buff *skb,
2783                        const struct nlmsghdr *nlh,
2784                        const struct nlattr * const cda[])
2785 {
2786         if (nlh->nlmsg_flags & NLM_F_DUMP) {
2787                 struct netlink_dump_control c = {
2788                         .dump = ctnetlink_exp_stat_cpu_dump,
2789                 };
2790                 return netlink_dump_start(ctnl, skb, nlh, &c);
2791         }
2792
2793         return 0;
2794 }
2795
2796 #ifdef CONFIG_NF_CONNTRACK_EVENTS
2797 static struct nf_ct_event_notifier ctnl_notifier = {
2798         .fcn = ctnetlink_conntrack_event,
2799 };
2800
2801 static struct nf_exp_event_notifier ctnl_notifier_exp = {
2802         .fcn = ctnetlink_expect_event,
2803 };
2804 #endif
2805
2806 static const struct nfnl_callback ctnl_cb[IPCTNL_MSG_MAX] = {
2807         [IPCTNL_MSG_CT_NEW]             = { .call = ctnetlink_new_conntrack,
2808                                             .attr_count = CTA_MAX,
2809                                             .policy = ct_nla_policy },
2810         [IPCTNL_MSG_CT_GET]             = { .call = ctnetlink_get_conntrack,
2811                                             .attr_count = CTA_MAX,
2812                                             .policy = ct_nla_policy },
2813         [IPCTNL_MSG_CT_DELETE]          = { .call = ctnetlink_del_conntrack,
2814                                             .attr_count = CTA_MAX,
2815                                             .policy = ct_nla_policy },
2816         [IPCTNL_MSG_CT_GET_CTRZERO]     = { .call = ctnetlink_get_conntrack,
2817                                             .attr_count = CTA_MAX,
2818                                             .policy = ct_nla_policy },
2819         [IPCTNL_MSG_CT_GET_STATS_CPU]   = { .call = ctnetlink_stat_ct_cpu },
2820         [IPCTNL_MSG_CT_GET_STATS]       = { .call = ctnetlink_stat_ct },
2821         [IPCTNL_MSG_CT_GET_DYING]       = { .call = ctnetlink_get_ct_dying },
2822         [IPCTNL_MSG_CT_GET_UNCONFIRMED] = { .call = ctnetlink_get_ct_unconfirmed },
2823 };
2824
2825 static const struct nfnl_callback ctnl_exp_cb[IPCTNL_MSG_EXP_MAX] = {
2826         [IPCTNL_MSG_EXP_GET]            = { .call = ctnetlink_get_expect,
2827                                             .attr_count = CTA_EXPECT_MAX,
2828                                             .policy = exp_nla_policy },
2829         [IPCTNL_MSG_EXP_NEW]            = { .call = ctnetlink_new_expect,
2830                                             .attr_count = CTA_EXPECT_MAX,
2831                                             .policy = exp_nla_policy },
2832         [IPCTNL_MSG_EXP_DELETE]         = { .call = ctnetlink_del_expect,
2833                                             .attr_count = CTA_EXPECT_MAX,
2834                                             .policy = exp_nla_policy },
2835         [IPCTNL_MSG_EXP_GET_STATS_CPU]  = { .call = ctnetlink_stat_exp_cpu },
2836 };
2837
2838 static const struct nfnetlink_subsystem ctnl_subsys = {
2839         .name                           = "conntrack",
2840         .subsys_id                      = NFNL_SUBSYS_CTNETLINK,
2841         .cb_count                       = IPCTNL_MSG_MAX,
2842         .cb                             = ctnl_cb,
2843 };
2844
2845 static const struct nfnetlink_subsystem ctnl_exp_subsys = {
2846         .name                           = "conntrack_expect",
2847         .subsys_id                      = NFNL_SUBSYS_CTNETLINK_EXP,
2848         .cb_count                       = IPCTNL_MSG_EXP_MAX,
2849         .cb                             = ctnl_exp_cb,
2850 };
2851
2852 MODULE_ALIAS("ip_conntrack_netlink");
2853 MODULE_ALIAS_NFNL_SUBSYS(NFNL_SUBSYS_CTNETLINK);
2854 MODULE_ALIAS_NFNL_SUBSYS(NFNL_SUBSYS_CTNETLINK_EXP);
2855
2856 static int __net_init ctnetlink_net_init(struct net *net)
2857 {
2858 #ifdef CONFIG_NF_CONNTRACK_EVENTS
2859         int ret;
2860
2861         ret = nf_conntrack_register_notifier(net, &ctnl_notifier);
2862         if (ret < 0) {
2863                 pr_err("ctnetlink_init: cannot register notifier.\n");
2864                 goto err_out;
2865         }
2866
2867         ret = nf_ct_expect_register_notifier(net, &ctnl_notifier_exp);
2868         if (ret < 0) {
2869                 pr_err("ctnetlink_init: cannot expect register notifier.\n");
2870                 goto err_unreg_notifier;
2871         }
2872 #endif
2873         return 0;
2874
2875 #ifdef CONFIG_NF_CONNTRACK_EVENTS
2876 err_unreg_notifier:
2877         nf_conntrack_unregister_notifier(net, &ctnl_notifier);
2878 err_out:
2879         return ret;
2880 #endif
2881 }
2882
2883 static void ctnetlink_net_exit(struct net *net)
2884 {
2885 #ifdef CONFIG_NF_CONNTRACK_EVENTS
2886         nf_ct_expect_unregister_notifier(net, &ctnl_notifier_exp);
2887         nf_conntrack_unregister_notifier(net, &ctnl_notifier);
2888 #endif
2889 }
2890
2891 static void __net_exit ctnetlink_net_exit_batch(struct list_head *net_exit_list)
2892 {
2893         struct net *net;
2894
2895         list_for_each_entry(net, net_exit_list, exit_list)
2896                 ctnetlink_net_exit(net);
2897 }
2898
2899 static struct pernet_operations ctnetlink_net_ops = {
2900         .init           = ctnetlink_net_init,
2901         .exit_batch     = ctnetlink_net_exit_batch,
2902 };
2903
2904 static int __init ctnetlink_init(void)
2905 {
2906         int ret;
2907
2908         pr_info("ctnetlink v%s: registering with nfnetlink.\n", version);
2909         ret = nfnetlink_subsys_register(&ctnl_subsys);
2910         if (ret < 0) {
2911                 pr_err("ctnetlink_init: cannot register with nfnetlink.\n");
2912                 goto err_out;
2913         }
2914
2915         ret = nfnetlink_subsys_register(&ctnl_exp_subsys);
2916         if (ret < 0) {
2917                 pr_err("ctnetlink_init: cannot register exp with nfnetlink.\n");
2918                 goto err_unreg_subsys;
2919         }
2920
2921         ret = register_pernet_subsys(&ctnetlink_net_ops);
2922         if (ret < 0) {
2923                 pr_err("ctnetlink_init: cannot register pernet operations\n");
2924                 goto err_unreg_exp_subsys;
2925         }
2926 #ifdef CONFIG_NETFILTER_NETLINK_QUEUE_CT
2927         /* setup interaction between nf_queue and nf_conntrack_netlink. */
2928         RCU_INIT_POINTER(nfq_ct_hook, &ctnetlink_nfqueue_hook);
2929 #endif
2930         return 0;
2931
2932 err_unreg_exp_subsys:
2933         nfnetlink_subsys_unregister(&ctnl_exp_subsys);
2934 err_unreg_subsys:
2935         nfnetlink_subsys_unregister(&ctnl_subsys);
2936 err_out:
2937         return ret;
2938 }
2939
2940 static void __exit ctnetlink_exit(void)
2941 {
2942         pr_info("ctnetlink: unregistering from nfnetlink.\n");
2943
2944         unregister_pernet_subsys(&ctnetlink_net_ops);
2945         nfnetlink_subsys_unregister(&ctnl_exp_subsys);
2946         nfnetlink_subsys_unregister(&ctnl_subsys);
2947 #ifdef CONFIG_NETFILTER_NETLINK_QUEUE_CT
2948         RCU_INIT_POINTER(nfq_ct_hook, NULL);
2949 #endif
2950 }
2951
2952 module_init(ctnetlink_init);
2953 module_exit(ctnetlink_exit);